A STRUCTURAL VIEW OF APPROXIMATIONsanjeev/papers/thesis.pdf · 1 Introduction 1 1.1 The Historical...

A STRUCTURAL VIEW OF APPROXIMATION

A DISSERTATION

SUBMITTED TO THE DEPARTMENT OF COMPUTER SCIENCE

AND THE COMMITTEE ON GRADUATE STUDIES

OF STANFORD UNIVERSITY

IN PARTIAL FULFILLMENT OF THE REQUIREMENTS

FOR THE DEGREE OF

DOCTOR OF PHILOSOPHY

By

Sanjeev Khanna

August 1996

c Copyright 1996 by Sanjeev Khanna

All Rights Reserved

ii

I certify that I have read this dissertation and that in my opinion it

is fully adequate, in scope and in quality, as a dissertation for the

degree of Doctor of Philosophy.

Rajeev Motwani(Principal Advisor)




Serge Plotkin




Madhu Sudan

Approved for the University Committee on Graduate Studies:

iii

AbstractThe discovery of efficient probabilistically checkable proofs for NP, and their surprising connection

to hardness of approximation, has resulted in a quantum leap in our understanding of approximability

of many important optimization problems. Yet much of this research has stayed focused on problem-

specific results. Building on many such results, this dissertation develops frameworks which unify a

variety of seemingly different results and thereby highlights the intrinsic structural properties which

govern the approximability of optimization problems.

Broadly speaking, our work comprises of three distinct parts. In the first part, we develop a

structural basis for computationally-defined approximation classes such as APX (constant-factor

approximable problems) and poly-APX (polynomial-factor approximable problems). We show that

there exist canonical transformations whereby every problem in an approximation class can be

“expressed” as a problem in a syntactically-defined optimization class. Thus, for example, we show

that every problem in APX is expressible as a problem in the syntactic class MAX SNP. These results

serve to reconcile two distinct views of approximability and provide a better understanding of the

structure of the computational classes. An immediate consequence of these results is the discovery

that a wide variety of natural optimization problems are complete for the computational classes.

We also study the possibility that the core of PTAS (problems with polynomial-time approximation

schemes) may be characterized via a combination of syntactic constraints and restrictions on the

input instances.

In the second part, we study two natural classes of maximization problems defined via constraint

satisfaction requirements. We discover classification theorems which, quite surprisingly, allow us

to determine the approximability of every problem in both these classes merely from the description

of the problem. Our results serve to formally affirm many trends concerning the approximability of

iv

optimization problems which have emerged over the years. In particular, they provide insight into

questions such as: Why are there no natural maximization problems that exhibit a log threshold

for approximability? Why do all the known MAX SNP problems that are NP-hard, turn out also to

be MAX SNP-hard? and, so on.

Finally, in the last part of this work, we study the problem of efficiently coloring a 3-colorable

graph. Since Karp’s original NP-hardness result for this problem, no better lower bounds have been

obtained towards the approximability of this fundamental optimization problem. We show that it

is NP-hard to color a 3-colorable graph with 4 colors and thereby establish that a -colorable graph

cannot be colored with less than 2 3 colors, unless P NP.

v

Acknowledgements

First and foremost, I would like to express my sincere thanks to my advisor Rajeev Motwani for his

continuous guidance and support. I have learned a lot from his creative style of research and his

persistence in tackling difficult research problems. It has been a privilege to work under his close

supervision.

Many thanks to Madhu Sudan who has been not only a partner in several of my research efforts

but also a friend, mentor and a constant source of inspiration.

Thanks also to Muli Safra for being a very encouraging and patient collaborator in my starting

days, and for teaching me many things in complexity theory.

I have had the great fortune of learning from many great teachers. In particular, I would like

to thank Ed Reingold, Doug West, Rajeev Motwani, and Serge Plotkin, for all the fascinating

algorithms, combinatorics, and graph theory courses they have taught me. I would also like to thank

them all for generously giving me their valuable advice on many occasions. Thanks are also due to

Serge for being on my thesis reading committee.

Very special thanks to Julien Basch and Chandra Chekuri who have been constant partners in my

academic and non-academic pursuits over the last several years. My life has been deeply enriched

by their wonderful friendship.

My stay at Stanford would not have been such a pleasant experience without the great com-

panionship of my colleagues. I would like to thank Anand, Arjun, Ashish, Craig, David, Donald,

Eric, Harish, Jeffrey, Luca, Michael, Moses, Omri, Piotr, Robert, Stevens, and Suresh, for being

delightful colleagues and for keeping me stimulating company over the years.

vi

I am truly indebted to my mother, father and brother, for their love, support and encouragement

from thousands of miles away.

Finally, I would like to thank my wonderful wife Delphine for always being there for me.

vii

Contents

Abstract iv

Acknowledgements vi

1 Introduction 1

1.1 The Historical Development 4

1.1.1 The Theory of NP-Completeness 4

1.1.2 The Advent of Approximation Algorithms 6

1.1.3 The Computational view of Approximation 8

1.1.4 Syntactic Characterizations of Decision Problems 9

1.1.5 Syntactic Optimization Classes 9

1.1.6 Probabilistically Checkable Proofs (PCP) 14

1.1.7 PCP and the Hardness of Approximation 15

1.2 Overview and Organization of Results 16

1.2.1 Part I: A Structural Basis for Approximation Classes 16

1.2.2 Part II: Approximability of the Constraint Satisfaction Problems 18

1.2.3 Part III: Hardness of Approximating Graph-Coloring 19

2 A Structure Theorem for Approximation Classes 21

2.1 Introduction 22

2.2 Organization of Results 25

2.3 E-Reductions 26

viii

2.4 MAX SNP Closure and APX-PB 28

2.4.1 Reducing Minimization Problems to Maximization Problems 29

2.4.2 NP Languages and MAX 3-SAT 30

2.4.3 Reducing Maximization Problems to MAX 3-SAT 31

2.5 Generic Reductions and an Approximation Hierarchy 33

2.5.1 The Reduction 34

2.5.2 Applications 38

2.6 Local Search and MAX SNP 40

2.6.1 Classical Local Search 40

2.6.2 Non-Oblivious Local Search 40

2.7 The Power of Non-Oblivious Local Search 42

2.7.1 Oblivious Local Search for MAX 2-SAT 42

2.7.2 Non-Oblivious Local Search for MAX 2-SAT 46

2.7.3 Generalization to MAX k-SAT 47

2.8 Local Search for CSP and MAX SNP 48

2.8.1 Constraint Satisfaction Problems 48

2.8.2 Non-Oblivious Local Search for MAX k-CSP 50

2.9 Non-Oblivious versus Oblivious GLO 52

2.9.1 MAX 2-CSP 52

2.9.2 Vertex Cover 53

2.10 The Traveling Salesman Problem 54

2.11 Maximum Independent Sets in Bounded Degree Graphs 55

3 Towards a Structural Characterization of PTAS 61

3.1 Introduction 62

3.2 Summary of Results and the Implications 64

3.2.1 Main Results and Implications 65

3.2.2 Possibility of Extension of Classes 68

3.2.3 Overview of Proof Techniques 69

ix

3.3 Outerplanar Graphs and Tree Decompositions 70

3.4 Planar MPSAT 72

3.5 Planar TMAX 76

3.6 Planar TMIN 78

3.7 1 -Outerplanar MINMAX 78

4 The Approximability of Constraint Maximization Problems 80

4.1 Introduction 81

4.2 Preliminaries 82

4.3 Overview of the Main Result 85

4.3.1 Discussion 86

4.3.2 Related Work 87

4.4 Polynomial Time Solvability 87

4.5 Proof of MAX SNP-Hardness 89

4.5.1 Notation 90

4.5.2 -Implementations and MAX SNP-Hard Functions 90

4.5.3 Characterizing 2-Monotone Functions 92

4.5.4 MAX SNP-hardness of Non 2-Monotone Functions 93

4.5.5 Implementing the REP Function 95

4.5.6 Implementing the Unary Functions 97

4.5.7 REP Helps Implement MAX SNP-Hard Functions 98

4.5.8 Unary Functions Help Implement either REP or MAX SNP-Hard Functions 98

4.6 Hardness at Gap Location 1 100

4.7 Strengthening Schaefer’s Dichotomy Theorem 101

5 The Approximability of Structure Maximization Problems 103

5.1 Introduction 104

5.2 Definitions 105

5.2.1 Problem Definition 105

5.2.2 Constraint Functions 106

x

5.2.3 Families of Constraint Functions 106

5.3 Overview of the Main Result 107

5.3.1 Main Theorem 107

5.3.2 Discussion 108

5.4 Preliminaries 108

5.4.1 Weighted vs. unweighted problems 109

5.4.2 Implementations and weighted problems 111

5.5 Proof of Main Theorem 114

5.5.1 Case 1: The Polynomial-Time Case 114

5.5.2 Case 4: The Decidable but Non-Approximable Case 115

5.5.3 Case 5: The NP-Hard Case 117

5.5.4 Case 2: The APX-Complete Case 117

5.5.4.1 Membership in APX 117

5.5.4.2 APX-Hardness 118

5.5.5 Case 3: The Poly-APX-Complete Case 121

5.5.5.1 Membership in poly-APX 121

5.5.5.2 poly-APX-hardness 121

6 The Approximability of Graph Coloring 129

6.1 Introduction 130

6.2 Hardness of Approximating the Chromatic Number when the Value May be Large 131

6.3 Hardness of Approximating the Chromatic Number when the Value is (a Large)

Constant 137

6.4 Hardness of Approximating the Chromatic Number for Small Values 140

6.4.1 The Structure of 140

6.4.2 A Clique of Size in Implies a 3-Coloring of ¯ 143

6.4.3 A 4-Coloring of ¯ Implies a Large Clique in 143

6.4.4 Off with the 4th Clique 153

7 Conclusion 154

xi

A Problem Definitions 157

A.1 SAT 157

A.2 -SAT 157

A.3 MAX -SAT 157

A.4 MAX CUT 158

A.5 - MIN CUT 158

A.6 MAX CLIQUE 158

A.7 MAX INDEPENDENT SET 158

A.8 GRAPH COLORING 158

A.9 GRAPH -COLORING 159

A.10 TSP(1,2) 159

A.11 MIN VERTEX COVER 159

A.12 MIN SET COVER 159

Bibliography 160

xii

Chapter 1

Introduction

1

CHAPTER 1. INTRODUCTION 2

Over the last three decades, a large number of naturally-arising optimization problems have been

shown to be NP-hard. Some classic examples of such problems include multiprocessor scheduling

problems arising in systems, constraint satisfaction problems arising in AI, graph coloring problems

arising in compiler optimization, and the well-known traveling salesman problem. Unless P NP,

there do not exist polynomial-time algorithms to solve these problems optimally. Despite the

intensive efforts of many researchers, there has been little progress towards solving the NP-hard

problems in polynomial-time. In fact, in the early 70’s itself, it was widely conjectured that P NP.

Since polynomial-time computability is commonly regarded as a pre-requisite in considering an

optimization problem as computationally tractable, researchers have considered relaxed versions of

these problems so as to make them provably tractable. One such well-studied relaxation has been to

give up the optimality condition, i.e., to allow approximate solutions. The standard measure used to

assess the quality of approximation is called the approximation ratio. Informally, an approximation

ratio measures the multiplicative factor which separates the optimal and the approximate solution.

In general, the objective is to find polynomial-time algorithms with approximation ratio close to

one.

Sustained research over the years has significantly enhanced our understanding of NP-hard

optimization problems from an algorithmic as well as a complexity-theoretic point of view. On

the one hand, we have polynomial-time algorithms for a variety of problems with provably good

approximation ratios. On the other hand, we now also have matching bounds, commonly referred

to as hardness of approximation results, which show that these ratios cannot be further improved

unless P NP (in which case, they can be in fact solved optimally). This research has brought to

the fore some surprising facts and intriguing trends. We mention some of the most important ones

along with the issues surrounding them. First, even though the decision problems underlying all

known NP-hard optimization problems are polynomially isomorphic, these optimization problems

themselves exhibit a wide spectrum of behavior with respect to polynomial-time approximability.

On the one end of this spectrum, there are problems such as multiprocessor scheduling which can be

approximated to within any constant strictly greater than one. On the other end of the spectrum, we

have problems such as graph coloring where we cannot approximate the optimal solution value to

within even certain polynomial factors. Is there an underlying structure which unifies all problems


with similar approximability? Second, “natural” optimization problems seem to exhibit only few

distinct types of approximability. For instance, we do not know of any natural maximization

problem which is approximable to within logarithmic factors and yet is not approximable to within

some constant factor. In fact, it seems consistent with our current understanding of approximability

to conjecture the non-existence of such problems. While there are several impediments to a

formal investigation of such a thesis in the general setting, can its truth be validated for any class of

natural problems which is well-structured and yet general enough to capture representative problems

from the entire spectrum of NP optimization problems? Third, the emergence of several widely

applicable algorithmic paradigms such as the primal-dual method [44] and the randomized rounding

technique [92] has highlighted that a suitable syntactic expressibility of a general class of problems

can be translated into a canonical positive approximability result. However, with the exception

of just one result1, there are no known canonical hardness results which apply to an entire class

of problems. Such results are critical to our understanding of intrinsic structure governing the

approximability of seemingly different problems. Are there other large classes of naturally-arising

optimization problems where syntactic expressibility can be used to uniformly distinguish “hard”

problems from “easy” ones? These are the issues which motivate the work of this dissertation.

Our work makes progress towards addressing each of the questions raised above. The next

two sections provide the necessary background and a formal overview of our results. However, a

brief informal description of some representative results follows. We show that all problems with

similar approximability can be expressed in a canonical manner. This canonical representation is

the intrinsic structural property shared by all of them. For example, we show that every problem

with a constant-factor approximation can be “expressed” as the MAX 3-SAT problem. We study

the approximability of maximization problems built out of constraint satisfaction requirements.

Our results yield a classification theorem whereby the precise asymptotic approximability of every

problem defined in this manner, can be determined merely from the problem description. The

classification theorem gives a finite approximation hierarchy for the problems in these classes and,

moreover, the levels of these hierarchies correspond precisely to those observed for naturally-arising

1This is the well-known result of Lund and Yannakakis where they studied hardness of approximating maximumsubgraph for hereditary graph properties [81].


optimization problems. This theorem formally affirms the trends which seem to emerge from the

study of optimization problems over the last two decades.

Much of the work in this dissertation builds upon the recent results which established a re-

markable connection between the existence of efficient probabilistically checkable proofs and the

approximability of optimization problems [37, 8, 7]. These results have led to a quantum leap in our

understanding of the hardness of approximation. Based on these largely problem-specific results,

our work has developed general results which apply to entire classes of optimization problems,

thereby underscoring the inherent structure governing the approximability of these optimization

problems.

1.1 The Historical Development

This section reviews the basic concepts through a historical overview of the results most intimately

tied to our work. The historical development presented here serves to set up the conceptual

foundations and the context for the work described in this dissertation.

1.1.1 The Theory of NP-Completeness

We start with a formal definition of the class NP; some preliminaries follow. An alphabet is a finite

set of symbols, commonly denoted by Σ. A string is a finite sequence of symbols drawn from an

alphabet. The length of a string is denoted by . A language is a set of strings drawn over a

fixed alphabet, i.e., Σ .

Definition 1 (NP) A language Σ belongs toNP if there exists a polynomial-time deterministic

Turing machine and a constant such that for all Σ , we have

Σ with such that accepts the input pair .

Σ , rejects the input pair .

A natural interpretation of the above definition is as follows: the string is the proof to the

assertion that and the machine is the verifier of this proof.


The theory of NP-completeness was originally developed to study decision problems. Cook [25]

and Levin [76] independently established that all NP-languages are “reducible” to the 3-SAT

problem2 — the problem of deciding if a given 3-SAT formula is satisfiable or not. Soon thereafter,

Karp [57] showed that the 3-SAT problem in turn, is “equivalent” to a whole variety of natural NP

decision problems. The notion of reduction and equivalence are formalized in the definitions below.

Definition 2 (Polynomial-Time Reduction) A language 1 Σ1 is polynomial-time reducible to

another language 2 Σ2 (denoted 1 2) if there exists a polynomial-time computable

function : Σ1 Σ2 such that for all Σ1, 1 if and only if 2.

This notion of reduction is called polynomial-time many-one reduction or Karp reduction in the

literature. A language 1 is polynomial-time equivalent to another language 2 if 1 2 and

2 1.

Definition 3 (NP-Hard) A language is NP-hard if for any language NP, .

Definition 4 (NP-Complete) A language is NP-complete if NP and is NP-hard.

The Cook-Levin result established that the 3-SAT problem is NP-complete. Karp showed that

many important NP decision problems such as determining whether a graph is -colorable, verifying

the hamiltonicity of a graph, or testing if a graph has a clique of size at least , are all reducible to

the 3-SAT problem and hence NP-hard. Thus, either all these problems are hard (i.e., P NP), or

they are all easy (i.e., P NP). Simultaneously, it was also realized that NP-completeness theory

provides a framework to study the complexity of many naturally-arising optimization problems.

These two developments had a remarkable impact in stimulating the theory of NP-completeness and

establishing the importance of the P NP question. The class of optimization problems studied

via the NP decision problems is called NPO, an acronym for NP optimization problems.

Definition 5 (NPO) An NPO problem Π is a four-tuple such that:

is the set of input instances and it is recognizable in polynomial-time,

2See Appendix A for a formal definition of various well-known problems that we will refer to in this dissertation.


Given any input , denotes the set of feasible solutions such that there exists a

polynomial with for all . Moreover, there exists a polynomial-time

computable predicate such that holds if and only if .

Given , denotes the objective function value and is polynomial-time com-

putable.

Finally, goal max min denotes whether Π is a maximization or a minimization problem.

The NPO problems are reduced to decision problems by considering the feasibility of solutions

which satisfy a given bound on the objective function value. The resulting decision problem is

at least as hard as the original optimization problem, and in fact often the complexity of the two

problems is related by only a polynomial factor. Over the years, more and more optimization

problems have been shown to have an associated NP-complete decision problem. Thus, NP-

completeness theory became a vehicle for establishing the intractability of NPO problems, modulo

the assumption P NP. But of course, this methodology required that the optimization problems

first give up their essential character of being optimization problems!

1.1.2 The Advent of Approximation Algorithms

All attempts to resolve the P NP question thus far have been in vain. In the early 70’s itself,

it began to be widely conjectured that these two classes are distinct and that it is unlikely that we

have polynomial-time exact algorithms for the NPO problems. Since many of these problems are

well-rooted in practice, computational tractability is an important objective. It was only natural that

researchers begin considering relaxation of these problems to obtain provably tractable versions

of these problems. The primary such relaxation considered was that of relaxing the optimality

condition. In particular, the focus shifted to polynomial time algorithms to compute near-optimal

solutions, i.e., solutions which are guaranteed to be within a certain multiplicative factor of the

optimal solution.

Definition 6 (Approximation Algorithm)An algorithm is an approximation algorithm for aNPO

problem Π if given an input instance , it computes a feasible solution for the input .


Of course, the value of an arbitrary feasible solution can be far from the optimal value. Our

interest is in algorithms which return solutions with guaranteed near-optimal values. The following

definitions help us characterize the performance of approximation algorithms. We use the notation

OPT to denote the optimum objective function value on instance of a problem. Throughout

this dissertation, we study only polynomial-time approximation algorithms.

Definition 7 (Performance Ratio) An approximation algorithm for an optimization problem Π

has performance ratio if, given any instance of Π with , the solution satisfies

maxOPT

OPT

A solution of value within a multiplicative factor of the optimal value is referred to as an

-approximation. We say that a NPO problem is approximable to within a factor if it has

a polynomial-time approximation algorithm with performance ratio .

In a seminal paper, Johnson [58] studied the polynomial-time approximability of many NPO

problems. He discovered that while good polynomial-time approximations are possible for several

NP-hard optimization problems, there are many others which seem to resist efforts towards such

algorithms. The subsequent work of several other researchers has only served to confirm the trend

observed by Johnson. We now have a whole variety of results which show that many NP-hard

optimization problems cannot be well-approximated unless P NP. More importantly, these

results have served to formally confirm the diverse approximability of the NP-hard optimization

problems. This diverse behavior might appear somewhat surprising at first glance because the

decision problems underlying these optimization problems were not only polynomially reducible to

each other but were also polynomially isomorphic3. But clearly, the reductions among the decision

problems were not strong enough to preserve the quality of approximate solutions. There was a

distinct need for developing a framework which allowed us to study optimization problemswithout

having to deprive them of their essential character by disguising them as decision problems.

3Two languages 1 2 Σ are said to be polynomially isomorphic if there is a polynomial-time computablebijection from Σ to itself such that 1 if and only if 2. Hartmanis and Berman [51] showed in 1976 thatall known NP-complete problems are polynomially isomorphic.


Meanwhile, a behavioral classification scheme emerged for the NPO problems — one which

recognized their differences in terms of approximability.

1.1.3 The Computational view of Approximation

This classification scheme was implicit in Johnson’s work — NPO problems are assigned to various

classes based on their polynomial-time approximability. Thus, we have classes such as APX and

PTAS.

Definition 8 (APX) An NPO problem Π is in the class APX if there exists a polynomial-time

algorithm for Π whose performance ratio is bounded by a constant.

Definition 9 (PTAS) An NPO problemΠ is in the class PTAS if for any rational 0, there exists

a polynomial-time approximation algorithm forΠ whose performance ratio is bounded by 1 .

If we let -APX denote the class of NPO problems that are approximable to within a factor ,

then we obtain a hierarchy of approximation classes. For instance, poly-APX and log-APX are the

classes of NPO problems which have polynomial-time algorithms with performance ratio bounded

polynomially and logarithmically, respectively, in the input length.

The strength of this classification scheme lies in its ability to capture all problems with a

specified threshold of approximability. But this view of approximation is an “abstract view”; it does

not highlight the inherent structural properties which determine the approximability of the NPO

problems.

We will often be interested in restricting our attention to polynomially bounded subsets of these

classes.

Definition 10 (Polynomially Bounded Subset) The polynomially bounded subset of class of

problems, denoted -PB, is the set of all problems Π for which there exists a polynomial

such that for all instances Π, OPT .

Thus, for example, APX-PB denotes the class of NPO problems that are constant-factor approx-

imable and have polynomially-bounded optimum values.


1.1.4 Syntactic Characterizations of Decision Problems

The notion of complexity seems inherently linked to computational models and resource bounds on

space and time. But surprisingly, almost parallel to the development of the computation-oriented

view of NP-completeness theory, logicians were successfully developing a descriptive or syntactic

view of complexity classes. Such efforts are rooted in a very simple intuition, namely, a “hard

to decide” problem is probably also “hard to describe”, and vice versa. A classical result in this

direction was established by Fagin [34].

Theorem 1 ([34]) NP is precisely the set of decision problems that can be expressed in the existential

second order logic, that is as Φ where Φ is a first-order formula, is a structure (i.e.

it is a second-order variable which ranges over relations of a fixed arity over the input universe)

whose existence is to be decided, and ; is the input structure comprising of a universe

and a finite set of constant arity predicates .

This is a remarkable theorem. A complexity class which was defined in terms of Turing machine

computations performed in a certain amount of time can also be described with no reference to either

machines or resources. This characterization clearly tells us the structure of the NPlanguages. Using

this characterization of NP, it is a relatively intuitive task to establish 3-SAT as NP-complete. The

ease of discovering natural complete problems is a very desirable attribute which seems readily

available with syntactically-defined classes, and becomes a difficult task (if not infeasible), when a

class is behaviorally-defined.

1.1.5 Syntactic Optimization Classes

Over the last two decades, pioneering work due to Immerman and others has led to such syntactic

characterizations of many other classes, e.g., PSPACE and NSPACE (see the introductory sur-

vey [55]). But this perspective stayed essentially in the realm of decision problems for a long time.

It was as recently as 1988 that Papadimitriou and Yannakakis in a seminal paper [87] brought this

view to the study of optimization problems. They were motivated by a collection of NPO problems

which were known to be in APX, but whose membership in PTAS remained unresolved. It was


natural to ask if the question of their membership in PTAS could be reduced to a single question

just as the P NP question could be reduced to the question of the membership of 3-SAT in P.

There were two fundamental obstacles in achieving such a goal. First, how do we define a

class which naturally contains all these problems and yet is restrictive enough to essentially capture

only the relevant problems? Secondly, what notion of reductions would allow us to establish

an equivalence between these problems in terms of their approximability? Papadimitriou and

Yannakakis observed that efforts to computationally define such a class are not likely to succeed.

In their own words: “The intuitive reason is that computation is an inherently unstable, non-robust

mathematical object, in the sense that it can be turned from non-accepting by changes that would

be insignificant in any reasonable metric — say, by flipping a single state to accepting .........

There seems to be no clear notion of "approximately correct computaion". If we move the focus of

approximability to, for example, the number of correct bits in the input (so that the machine accepts),

then there seems to be no generic reduction that preserves approximability. ” Thus intuitively it

seemed clear that one needed to take a different approach towards defining such a class. The only

other such approach known was the syntactic approach hitherto applied to characterizing classes of

decision problems. Indeed, they discovered an elegant answer using the syntactic approach.

Papadimitriou and Yannakakis began with Fagin’s syntactic characterization of NP. In a

standardized form, Fagin’s result may be stated as follows: A decision problem belongs to the class

NP if and only if it can be expressed as where is a quantifier-free first

order formula, is the desired structure, ; is the input structure, and and are finite

vectors (i.e. they range over tuples in the universe of the input structure). Let us better understand

this definition through a concrete example.

Example 1 (SAT) The structure that we seek here is a truth assignment and we would like to verify

that it satisfies all the given clauses. The input consists of a universe whose elements are the

clauses and the variables, and a set of three predicates and . The unary predicate is true

if and only if a given element is a clause. The other two predicates and are binary, where

is true if and only if the clause contains the variable as a positive literal, and is

true if and only if the clause contains the variable as a negative literal. Now, the satisfiability


problem may be expressed as ; where

;

A natural subclass of NP, called strictNP or SNP, can be defined from the above characterization

simply by dropping the existential quantifier, that is, SNP is the class of properties expressible as

. A canonical example of a problem in this subclass is 3-SAT.

Example 2 (3-SAT) As before, the structure that we seek here is a truth assignment and we would

like to verify that it satisfies all the given clauses. The role of the existential quantifier in the

previous example is simply to verify that at least one literal in each clause is set to true. But in

3-SAT, the clauses have at most three literals each and hence the role of the existential quantifier

can be pushed into the fixed-size formula . More formally, we encode the input instance via four

predicates 0 1 2 and 3, where evaluates to true on a clause if and only if it has exactly

negated literals. Thus, if the clause 1 2 3 , then the first variables in appear

negated and the remaining variables appear unnegated. The 3-SAT problem may now be expressed

as 1 2 3 ; 0 1 2 3 1 2 3 , where

; 0 1 2 3 1 2 3 1 2 3 0 1 2 3

1 2 3 1 1 2 3

1 2 3 2 1 2 3

1 2 3 3 1 2 3

Now, what if we relax the condition that the structure must satisfy the formula for all choices

of ? That is, we look for structures which maximize the number of choices of for which the

formula is satisfied. This gives us an optimization analog of the classes SNP and NP.

Definition 11 (MAX SNP) An optimization problem Π is in MAX SNP if it can be expressed as


the problem of finding a structure which maximizes the objective function

:

where is the input and is a quantifier-free first order formula.

Example 3 (MAX 3-SAT) The formula is the same as in Example 2. But the goal is now to

find such that 1 2 3 : ; 0 1 2 3 1 2 3 is maximized.

A natural extension is to associate a weight with every tuple in the range of the universal

quantifier; the modified objective is to find an which maximizes Φ ,

where denotes the weight associated with the tuple .

The optimization analog of the class NP is the class MAX NP as defined below.

Definition 12 (MAX NP) An optimization problem Π is in MAX NP if it can be expressed as the

problem of finding a structure which maximizes the objective function

:

where is the input and is a quantifier-free first order formula.

A typical example of a problem that belongs to the class MAX NP but not to MAX SNP, is the

problem MAX SAT.

An immediate consequence of the syntactic definitions of these classes is a canonical approxima-

tion algorithm yielding a constant-factor approximation to every problem in both these classes [87].

Thus, both these classes are contained in the approximation class APX. However, it is not difficult

to show (via syntactic considerations) that many problems in APX are not expressible as a problem

in either of these classes, and hence that the containment is strict (see [74], for example).

Soon after the work of Papadimitriou and Yannakakis, a variety of other syntactic optimization

classes were introduced and well-studied in the literature. We formally introduce some of these

classes, such as RMAX [85] and MIN F Π2 [73], in the next chapter. In the remainder of this


section, we focus on the class MAX SNP. We first describe the notion of approximation-preserving

reduction used by Papadimitriou and Yannakakis; we begin with the concept of error in a solution.

Definition 13 (Error) Given a solution to an instance of anNPO problemΠ, we define its error

as

maxOPT

OPT1

Notice that the above definition of error applies uniformly to the minimizationand maximization

problems at all levels of approximability.

Definition 14 ( -Reduction) A problem Π -reduces to another problem Π (denoted Π Π )

if there exist polynomial-time computable functions and , and two positive constants and ,

satisfying the following:

maps an instance of Π to an instance of Π such that OPT OPT ,

maps solutions of to solutions of such that

OPT OPT

It is easy to verify the following two facts [87].

Fact 1 -reductions compose together, that is, ifΠ Π and Π Π , then Π Π .

Fact 2 -reductions preserve approximability. If Π Π , then given a solution to Π with

error , we can obtain a solution to Π with error at most .

This linear relation between the two errors gives the -reductions their name — linear reduc-

tions.

Akin to the notion of NP-hardness, Papadimitriou and Yannakakis defined the notion of

MAX SNP-hardness using -reductions4.

4Later, we slightly modify the notion of MAX SNP-hardness by introducing somewhat more general reductions.


Definition 15 (MAX SNP-Hard) An optimization problem Π is said to be MAX SNP-hard if for

any Π MAX SNP, Π Π.

Definition 16 (MAX SNP-Complete) An optimization problem Π is MAX SNP-complete if Π

MAX SNP and Π is MAX SNP-hard.

The approximation-preserving nature of -reductions ensures that if any MAX SNP-hard prob-

lem has a PTAS, then so does every problem in the entire class MAX SNP. A wide variety of

important optimization problems has been shown to be MAX SNP-hard. Some examples include

MAX 3-SAT, MIN VERTEX COVER, TSP(1,2), MAX INDEPENDENT SET in bounded degree

graphs (MIS-B), and so on. A positive or a negative approximability for any of these problems

would imply the same for the rest of the class — a development very similar to the unification

facilitated by the theory of NP-completeness.

1.1.6 Probabilistically Checkable Proofs (PCP)

As we saw earlier, there is a fundamental connection between verification of proofs and the definition

of the complexity class NP. We essentially defined NP as the class of languages whose membership

proofs are efficiently verifiable. The notion of efficient verification is based on polynomial-time

verification in the classical definition where the verifier examines the complete proof and the

verification process is assumed to be deterministic. In the mid-80’s, however, a new notion of proof

verification emerged. In this notion, the verifier is allowed an access to a random source and it

verifies the proof by examining only some of its bits. Thus, in the new framework, a verifier need

not read the complete proof. Of course, a consequence is that the verifier can no longer be certain

of having made a correct decision. The efficiency of verification is characterized by the number

of random bits needed and the query bits read by the verifier. This gave rise to a new family of

complexity classes as defined below.

Definition 17 (PCP) Given two functions : , a language is said to be in PCP

if there exists a probabilistic polynomial-time oracle machine such that for all Σ , we have

such that on all choices of random strings, on oracle accepts .


, on oracle accepts with probability at most 1/2.

Moreover, uses coin flips and queries the oracle at most times.

The above definition was formalized by Arora and Safra [8], and has its origins in the work

of Babai [11] and Goldwasser, Micali and Rackoff ([46]). The classical definition of NP can be

restated in terms of the above definition: NP PCP 0 where denotes the class

of univariate polynomials. However, an alternate surprising characterization of NP was provided

by Arora, Lund, Motwani, Sudan and Szegedy [7] (a detailed exposition of this result can be found

in [2, 97]).

Theorem 2 ([7]) NP PCP log 1 .

This result says that if there exists a polynomial size membership proof, then there also exists

another polynomial size membership proof which can be verified with high probability by just

examining a constant number of its bits! As we are going to see in the next section, PCP based

characterizations of NP has far-reaching ramifications for the hardness of approximation.

1.1.7 PCP and the Hardness of Approximation

The class MAX SNP, as defined by Papadimitriou and Yannakakis, gave the right equivalence class

to address the question of whether a PTAS indeed existed for many constant-factor approximable

optimization problems. This unification brought by this class is analogous to Karp’s work showing

equivalence between a whole variety of NP decision problems. But we still needed an analog of

Cook’s theorem for the class MAX SNP, that is, a result which would associate the question of

PTAS for MAX SNP-hard problem with the collapse of two complexity classes into one. This

critical final result came from PCP, a culmination of a long series of remarkable results.

In a pioneering paper, Feige et al [37] established a very surprising connection between ef-

ficiently checkable proofs and hardness of approximation. Specifically, they showed that if

NP PCP then there does not exist a constant-factor approximation algorithm for clique

unless NP DTIME 2q r . Soon thereafter, a connection to MAX 3-SAT was also realized,


namely, MAX 3-SAT does not have a PTAS unless P NP [7]. In one swoop, this final result

excludes the possibility of PTAS for any MAX SNP-hard problem unless P NP.

The sequence of events highlighting the deep connection between probabilistically checkable

proofs and the hardness of approximation may appear surprising at first. But, in hindsight, one

sees the intuition behind such a connection — the notion of probabilistic verification overcomes

the bottleneck mentioned by Papadimitriou and Yannakakis ([87]), namely, the inherently unstable

and non-robust nature of computation. Thus, just as the notion of exact verification in the classical

definition of NP provided a means to study the intractability of computing exact solutions to the

NP-hard optimization problems, the notion of probabilistic checking provided a means to study the

intractability of computing approximate solutions to the NP-hard optimization problems.

1.2 Overview and Organization of Results

Broadly speaking, this dissertation can be viewed as being composed of three distinct parts.

1.2.1 Part I: A Structural Basis for Approximation Classes

As pointed out earlier, the classification of NPO problems into classes such as APX and poly-APX,

does not yield insight into the structure of problems contained in these classes. We seek a charac-

terization for the approximation classes akin to Fagin’s syntactic characterization of NP which, in a

similar manner, identifies the structural properties shared by all problems in each of these classes. A

simple analogy from material sciences may help clearly illustrate the spirit of our quest. Materials

can be classified as soft, hard, or brittle, based on their observed toughness. But alternately, one

could associate the strength of the material with a certain underlying crystal structure, and thus obtain

a structural characterization for all materials that are soft, hard, or brittle. Computationally-defined

approximation classes are analogous to the classification of materials into soft, hard, or brittle, and

what we seek is an understanding of the structure that underlies their observed approximability.

Our work in this part shows that indeed approximation classes have such syntactic character-

izations. We show that with each of the computational classes, such as APX-PB, log-APX-PB

and poly-APX-PB, we can associate a syntactic optimization class such that every problem in


the computational class is “expressible” via reductions as a problem in the associated syntac-

tic class. For instance, we show that every problem in APX has an approximation-preserving

reduction to a problem in MAX SNP. Our method introduces a general technique for creat-

ing approximation-preserving reductions which show that any well-approximable problem can be

reduced in an approximation-preserving manner to a problem that is hard to approximate to corre-

sponding factors. We demonstrate this technique by applying it to other syntactic classes. Thus we

establish that the syntactic optimization classes provide a structural basis for the computationally-

defined approximation classes.

This reconciliation of the computational and the syntactic views enhances our understanding of

the approximation classes from both an algorithmic and a complexity-theoretic perspective. On the

one hand, we can now identify natural complete problems for classes such as APX, while on the

other hand, we can study algorithmic paradigms which uniformly apply to the set of core problems

in the approximation classes.

Chapter 2 describes most of the aforementioned work. In the later half of this chapter, we

study an algorithmic paradigm for MAX SNP. We use the syntactic nature of MAX SNP to

define a general paradigm, non-oblivious local search, useful for developing simple yet efficient

approximation algorithms. We show that such algorithms can find good approximations for all

MAX SNP problems, yielding approximation ratios comparable to the best-known for a variety

of specific MAX SNP-hard problems. Non-oblivious local search provably out-performs standard

local search in both the degree of approximation achieved and the efficiency of the resulting

algorithms.

While the techniques discussed in Chapter 2 lead to a syntactic characterization of several

approximation classes, they do not extend to the class PTAS. In fact, we are not aware of any natural

syntactic optimization classes which are subclasses of PTAS. Chapter 3 explores the possibility

that a core of PTAS may be characterized through syntactic classes endowed with restrictions on

the structure of the input instances.

We argue that while the core of APX is the purely syntactic class MAX SNP, in the case of PTAS

we must identify the core in terms of syntactic prescriptions for the problem definition augmented

with structural restrictions on the input instances. Specifically, we propose such a unified framework


based on syntactic classes restricted to instances exhibiting a planar structure. A variety of known

results, and some new results, follow directly from this framework, thereby lending credence to our

hypothesis that there exists some common structural underpinnings for problems in PTAS.

1.2.2 Part II: Approximability of the Constraint Satisfaction Problems

In this part of the dissertation, we initiate a systematic study of the maximization versions of

constraint satisfaction problems. On the one hand, our investigation leads to some surprising clas-

sification theorems which allow us to uniformly identify hard problems in these classes. More

precisely, we show that the syntactic description of these classes can be used for an exact charac-

terization of the central elements that govern the approximability of a problem in the class. On the

other hand, this study also serves to formally confirm some trends in approximability which have

emerged over the past several years.

The starting point in our study is the remarkable paper of Schaefer [95] which considers a

subclass of languages in NP and proves a “dichotomy theorem” for this class. The subclass

considered was problems expressible as “constraint satisfaction problems” and the dichotomy

theorem showed that every language in this class is either in P, or is NP-hard. This result is in sharp

contrast to a result of Ladner [75], which shows that such a dichotomy does not hold for NP, unless

P NP.

We consider two classes of maximization problems built on Schaefer’s framework of constraint

satisfaction problems: the constraints are boolean constraints as in Schaefer’s case, but the objective

in one class is to find a solution maximizing the weight of satisfied constraints, while in the other,

we seek a feasible solution which maximizes the weight of variables set to true. Together, these

classes capture a whole variety of natural optimization problems such as MAX CUT, - MIN CUT

in directed graphs, and MAX CLIQUE. We determine the approximability of every problem in

each of these classes.

For the first class of problems, studied in Chapter 4, we show that every problem is either

solvable exactly in P or is APX-hard (hence not approximable to within arbitrary constant factors in

polynomial-time, unless P NP). The class of problems captured here is a subclass of MAX SNP.

We feel that this class forms a combinatorial core of MAX SNP and our results for this class bring


a new insight towards understanding the class MAX SNP.

The problems in the second class, on the other hand, are shown to belong to one of five classes:

exactly solvable in polynomial-time, approximable to within constant factors in polynomial-time,

approximable to within polynomial factors in polynomial-time, not approximable but decidable,

and undecidable (unless P=NP). Thus, this class has a representative problem at each level of ap-

proximability threshold observed for maximization problems. Our results here give some evidence

towards the truth of approximability patterns that have emerged from our study of maximization

problems. This latter class of problems is studied in Chapter 5.

Our work here builds on the impressive collection of non-approximability results derived over

the recent years (see the surveys [6, 12, 15, 59], for example). However, much of the past work has

been enumerative in its approach, considering specific problems and then proving hardness results

for such problems. Our approach adds to this body of work by extracting exhaustive results from

these — that is, we consider large (infinite) collections of problems and then characterize precisely

the asymptotic approximability of each one of these problems.

1.2.3 Part III: Hardness of Approximating Graph-Coloring

The last five years have witnessed an astounding progress in our understanding of the approx-

imability of NPO problems. An extensive collection of open problems has been either solved

completely or a significant progress has been made towards the final answer. The ( 1 )-hardness

results for MAX CLIQUE [52] and COLORING [38], the log -threshold for approximating

MIN SET COVER [36], the PTAS for the Euclidean TSP problem [3], the significantly improved

approximation ratios for MAX CUT [45], MAX 2-SAT [45], and GRAPH -COLORING [65], are

only some among this impressive collection of results. Yet there are important questions concerning

the approximability of fundamental optimization problems that remain unanswered. Probably the

most striking of these open questions is the approximability of the GRAPH 3-COLORING. Since

Karp’s original NP-hardness result over two decades ago, nothing more has been established on

the approximability of this problem. The best known approximation algorithms has a polynomial

approximation ratio [65]. In Chapter 6, we study this problem and make a modest progress (but

the only progress thus far), in improving the lower bound for this problem. Specifically, we show


that coloring a 3-colorable graph with 4 colors is NP-hard. This modest progress already requires

somewhat involved ideas. An immediate corollary of this result is that a -colorable graph cannot

be colored with less than 2 3 colors unless P NP. We also present a simplified proof

of the Ω lower bound due to Lund and Yannakakis [82] on the approximability of the general

graph coloring problem.

Chapter 2

A Structure Theorem for

Approximation Classes

21

CHAPTER 2. A STRUCTURE THEOREM FOR APPROXIMATION CLASSES 22

2.1 Introduction

We saw earlier that a variety of classes of NPO problems have been defined based on their

polynomial-time approximability. Some examples of these classes are APX (the class of constant-

factor approximable problems), PTAS (the class of problems with polynomial-time approximation

schemes), and FPTAS (the class of problems with fully-polynomial-time approximation schemes).

The advantage of working with classes defined using approximability as the criterion is that it allows

us to work with problems whose approximability is well-understood. Crescenzi and Panconesi [28]

have recently also been able to exhibit complete problems for such classes, particularly APX. Un-

fortunately such complete problems seem to be rare and artificial, and do not seem to provide insight

into the more natural problems in the class. Research in this direction has to find approximation-

preserving reductions from the known complete but artificial problems in such classes to the natural

problems therein, with a view to understanding the approximability of the latter.

We also saw that over the last decade, a second family of classes of NPO problems has

emerged.These are the classes defined via syntactic considerations, based on a syntactic charac-

terization of NP due to Fagin [34]. Research in this direction, initiated by Papadimitriou and

Yannakakis [87], and followed by Panconesi and Ranjan [85] and Kolaitis and Thakur [73], has led

to the identification of classes such as MAX SNP, RMAX(2), and MIN F Π2 1 . The syntactic

prescription in the definition of these classes has proved very useful in the establishment of complete

problems. Moreover, the recent results of Arora, Lund, Motwani, Sudan, and Szegedy [7] have

established the hardness of approximating complete problems for MAX SNP to within (specific)

constant factors unless P NP. It is natural to wonder why the hardest problems in this syntactic

sub-class of APX should bear any relation to all of NP.

Though the computational view allows us to precisely classify the problems based on their

approximability, it does not yield structural insights into natural questions such as: Why certain

problems are easier to approximate than some others? What is intrinsic to optimization problems

with a similar approximability behaviour? What is the canonical structure of the hardest represen-

tative problems of a given approximation class? and, so on. Furthermore, intuitively speaking, this

This chapter is based on joint work with Rajeev Motwani, Madhu Sudan and Umesh Vazirani [67].


view is too abstract to facilitate the identification of, and reductions to establish, natural complete

problems for a class. Showing existence of natural complete problems is a critical step in establish-

ing the importance of a complexity class. The importance of the class NP, for instance, is strongly

linked to its abundance in naturally arising important problems. The syntactic view, on the other

hand, is essentially a structural view. The syntactic prescription gives a natural way of identify-

ing canonical hard problems in the class and performing approximation-preserving reductions to

establish complete problems.

Attempts at trying to find a class with both the above mentioned properties, i.e., natural complete

problems and capturing all problems of a specified approximability, have not been very successful.

Typically the focus has been to relax the syntactic criteria to allow for a wider class of problems to

be included in the class. However in all such cases it seems inevitable that these classes cannot be

expressive enough to encompass all problems with a given approximability. This is because each

of these syntactically defined approximation classes is strictly contained in the class NPO; the strict

containment can be shown by syntactic considerations alone. As a result if we could show that any

of these classes contains all of P, then we would have separated P from NP. We would expect that

every class of this nature would be missing some problems from P, and this has indeed been the

case with all current definitions.

We explore a different direction by studying the structure of the syntactically defined classes

when we look at their closure under approximation-preserving reductions. The idea of looking

at the closure of a class is implicit in the work of Papadimitriou and Yannakakis [87] who state

that: minimization problems will be “placed” in the classes through -reductions to maximization

problems. The advantage of looking at the closure of a set is that it maintains the complete problems

of the set, while managing to include all of P into the closure (for problems in P, the reduction is to

simply use a polynomial time algorithm to compute an exact solution). It now becomes interesting,

for example, to compare the closure of MAX SNP (denoted MAX SNP) with APX. A positive

resolution, i.e., MAX SNP APX, would immediately imply the non-existence of a PTAS for

MAX SNP-hard problems, since it is known that PTAS is a strict subset of APX, if P NP. On

the other hand, an unconditional negative result would be difficult to obtain, since it would imply

P NP.


Here we resolve this question in the affirmative. The exact nature of the result obtained depends

upon the precise notion of an approximation preserving reduction used to define the closure of

the class MAX SNP. The strictest notion of such reductions available in the literature are the

-reductions due to Papadimitriou and Yannakakis [87]. We work with a slight extension of the

reduction, which we call -reductions. Using such reductions to define the class MAX SNP we

show that this equals APX-PB, the class of all polynomially bounded NP optimization problems

which are approximable to within constant factors. An interesting side-effect of our results is the

positive answer to the question of Papadimitriou and Yannakakis [87] about whether MAX NP has

any complete problems. We next extend the technique used in showing this result to establish a

general structure theorem which shows that any "well" approximable problem can be reduced in

an approximation-preserving manner to a problem which is hard to approximate to corresponding

factors. The structure theorem provides a powerful tool for establishing that syntactic optimization

classes provide a structural basis for the computationally defined approximation classes. We

demonstrate this technique by applying it to other syntactic classes.

It is worthwhile to note that by using slightly looser definitions of approximation preserving

reductions (and in particular the PTAS-reductions of Crescenzi and Trevisan [29]), our results can

be extended to show that all of APX equals MAX SNP.

The syntactic view seems useful not only in obtaining structural complexity results but also in

developing paradigms for designing efficient approximation algorithms. This was demonstrated

first by Papadimitriou and Yannakakis [87] who show approximation algorithms for every problem

in MAX SNP. We further exploit the syntactic nature of MAX SNP to develop another paradigm

for designing good approximation algorithms for problems in that class and thereby provide an

alternate computational view of it. We refer to this paradigm as non-oblivious local search, and it is

a modification of the standard local search technique [99]. We show that every MAX SNP problem

can be approximated to within constant factors by such algorithms. It turns out that the performance

of non-oblivious local search is comparable to that of the best-known approximation algorithms for

several interesting and representative problems in MAX SNP. An intriguing possibility is that this

is not a coincidence, but rather a hint at the universality of the paradigm or some variant thereof.

Our results are related to some extent to those of Ausiello and Protasi [10]. They define a class


GLO (for Guaranteed Local Optima) of NPO problems which have the property that for all locally

optimum solutions, the ratio between the value of the global and the local optimum is bounded

by a constant. It follows that GLO is a subset of APX, and it was shown that it is in fact a strict

subset. We show that a MAX SNP problem is not contained in GLO, thereby establishing that

MAX SNP is not contained in GLO. This contrasts with our notion of non-oblivious local search

which is guaranteed to provide constant factor approximations for all problems in MAX SNP. In

fact, our results indicate that non-oblivious local search is significantly more powerful than standard

local search in that it delivers strictly better constant ratios, and also will provide constant factor

approximations to problems not in GLO. Independently of our work, Alimonti [1] has used a

similar local search technique for the approximation of a specific problem not contained in GLO or

MAX SNP.

2.2 Organization of Results

In Section 2.3, we introduce -reductions and the notion of closure of a class. In Section 2.4,

we show that MAX SNP APX-PB. A generic theorem which allows to equate the closure of

syntactic classes to appropriate computational classes is outlined in Section 2.5; we also develop an

approximation hierarchy based on this result.

The notion of non-oblivious local search and NON-OBLIVIOUS GLO is developed in Section 2.6.

In Section 2.7, we illustrate the power of non-obliviousness by first showing that oblivious local

search can achieve at most the performance ratio 3 2 for MAX 2-SAT, even if it is allowed to

search exponentially large neighborhoods; in contrast, a very simple non-oblivious local search

algorithm achieves a performance ratio of 4 3. We then establish that this paradigm yields a

2 2 1 approximation to MAX k-SAT. In Section 2.8, we provide an alternate characterization

of MAX SNP via a class of problems called MAX k-CSP. It is shown that a simple non-oblivious

algorithm achieves the best-known approximation for this problem, thereby providing a uniform

approximation for all of MAX SNP. In Section 2.9, we further illustrate the power of this class of

algorithms by showing that it can achieve the best-known ratio for a specific MAX SNP problem

and for VERTEX COVER (which is not contained in GLO). This implies that MAX SNP is not


contained in GLO, and that GLO is strict subset of NON-OBLIVIOUS GLO. In Section 2.10, we

apply it to approximating the traveling salesman problem. Finally, in Section 2.11, we apply this

technique to improving a long-standing approximation bound for maximum independent sets in

bounded-degree graphs.

2.3 E-Reductions

Given an NPO problem Π and an instance of Π, we use to denote the length of and OPT

to denote the optimum value for this instance. Also, recall that for any solution to , the value

of the solution is denoted by and is assumed to be a polynomial time computable function

which takes positive integer values (see Chapter 1 for formal definitions).

We now describe precisely the approximation preserving reduction that we will use. This

reduction, which we call the -reduction, is essentially the same as the -reduction of Papadimitriou

and Yannakakis [87] and differs from it in only one relatively minor aspect.

Definition 18 (E-reduction) A problem Π -reduces to a problem Π (denoted Π Π ) if there

exist polynomial time computable functions , and a constant such that

maps an instance ofΠ to an instance ofΠ such thatOPT andOPT are related by

a polynomial factor i.e. there exists a polynomial such that OPT OPT .

maps solutions of to solutions of such that

Remark 1 Among the many approximation preserving reductions in the literature, the -reduction

appears to be the strictest. The -reduction appears to be slightly weaker (in that it allows

polynomial scaling of the problems), but is stricter than any of the other known reductions. Since

all the reductions given in chapter are -reductions, they would also qualify as approximation-

preserving reductions under most other definitions and in particular, they fit the definitions of

-reductions and -reductions of Crescenzi and Panconesi [28].


Remark 2 HavingΠ Π implies thatΠ is as well approximable as Π ; in fact, an -reduction

is an FPTAS-preserving reduction. An important benefit is that this reduction can be applied

uniformly at all levels of approximability. This is not the case with the other existing definitions

of FPTAS-preserving reduction in the literature. For example, the FPTAS-preserving reduction

( -reduction) of Crescenzi and Panconesi [28] is much more unrestricted in scope and does not

share this important property of the -reduction. Note that Crescenzi and Panconesi [28] showed

that there exists a problemΠ PTAS such that for any problem Π APX, Π Π . Thus, there

is the undesirable situation that a problem Π with no PTAS has a FPTAS-preserving reduction to

a problem Π with a PTAS.

Remark 3 The -reduction of Papadimitriou and Yannakakis [87] enforces the condition that the

optima of an instance of Π be linearly related to the optima of the instance of Π to which

it is mapped. This appears to be an unnatural restriction considering that the reduction itself is

allowed to be an arbitrary polynomial time computation. This is the only real difference between

their -reduction and our -reduction, and an -reduction in which the linearity relation of the

optimas is satisfied is an -reduction. Intuitively, however, in the study of approximability the

desirable attribute is simply that the errors in the corresponding solutions are closely (linearly)

related. The somewhat artificial requirement of a linear relation between the optimum values

precludes reductions between problems which are related to each other by some scaling factor.

For instance, it seems desirable that two problems whose objective functions are simply related

by any fixed polynomial factor should be inter-reducible under any reasonable definition of an

approximation-preserving reduction. Our relaxation of the -reduction constraint is motivated

precisely by this consideration.

Let be any class of NPO problems. Using the notion of an -reduction, we define hardness and

completeness of problems with respect , as well its closure and polynomially-bounded sub-class.

Definition 19 (Hard and Complete Problems) A problem Π is said to be -hard if for all prob-

lems Π , we have Π Π . A -hard problem Π is said to be -complete if in addition

Π .


Definition 20 (Closure) The closure of , denoted by , is the set of allNPO problemsΠ such that

Π Π for some Π .

Remark 4 The closure operation maintains the set of complete problems for a class.

2.4 MAX SNP Closure and APX-PB

In this section, we will establish the following theorem and examine its implications. The proof is

based on the results of Arora et al [7] on efficient proof verifications.

Theorem 3 MAX SNP APX-PB.

Remark 5 The seeming weakness thatMAX SNP only captures polynomially boundedAPX prob-

lems can be removed by using looser forms of approximation-preserving reduction in defining the

closure. In particular, Crescenzi and Trevisan [29] define the notion of a PTAS-preserving reduc-

tion under which APX APX-PB. Using their result in conjunction with the above theorem, it

is easily seen that MAX SNP APX. This weaker reduction is necessary to allow for reductions

from fine-grained optimization problems to coarser (polynomially-bounded) optimization problems

(cf. [29]).

The following is a surprising consequence of Theorem 3.

Theorem 4 MAX NP MAX SNP.

Papadimitriou and Yannakakis [87] (implicitly) introduced both these closure classes but did

not conjecture them to be the same. It would be interesting to see if this equality can be shown

independent of the result of Arora et al [7]. We also obtain the following resolution to the problem

posed by Papadimitriou and Yannakakis [87] of finding complete problems for MAX NP.

Theorem 5 MAX SAT is complete for MAX NP.

The following sub-sections establish that MAX SNP APX-PB. The idea is to first -

reduce any minimization problem in APX-PB to a maximization problem in therein, and then


-reduce any maximization problem in APX-PB to a specific complete problem for MAX SNP,

viz., MAX 3-SAT. Since an -reduction forces the optimas of the two problems involved to be

related by polynomial factors, it is easy to see that MAX SNP APX-PB. Combining these two

facts, we obtain Theorem 3.

2.4.1 Reducing Minimization Problems to Maximization Problems

Observe that the fact that Π belongs to APX implies the existence of an approximation algorithm

and a constant such that

OPT OPT

Henceforth, we will use to denote . We first reduce any minimization problem

Π APX-PB to a maximization problem Π APX-PB, where the latter is obtained by merely

modifying the objective function for Π, as follows. Let Π have the objective function

max 1 1

for all instances and solutions for Π. Clearly, takes only positive values. To ensure

that is integer-valued, we can assume without loss of generality, that is an integer (a

real-valued performance ratio can always be rounded up to the next integer). It can be verified that

the optimum value for any instance of Π always lies between and 1 . Thus is a

1 -approximation algorithm for Π .

Now given a solution for instance of Π such that it has error , we want to construct a

solution for instance of Π such that the error is at most for some . We will show this for

1 .

First consider the case when 1 i.e. 1. In this case, we simply output

the solution . If 1 then we are trivially done else we observe that

1 1 1


On the other hand, if 1, we may proceed as follows. If is a -error solution to

the optimum of Π , i.e.,

OPT1

1

where OPT is the optimal value of for , we can conclude that

1

1 OPT OPT

OPT OPT

OPT 1 OPT

Thus a solution to Π with error is a solution to Π with error at most 1 , implying an

-reduction with 1.

2.4.2 NP Languages and MAX 3-SAT

The following theorem, adapted from a result of Arora, Lund, Motwani, Sudan, and Szegedy [7], is

critical to our -reduction of maximization problems to MAX 3-SAT.

Theorem 6 ([7]) Given a language NP and an instance Σ , one can compute in polynomial

time an instance of MAX 3-SAT, with the following properties.

1. The formula has clauses, where depends only on .

2. There exists a constant 0, independent of the input , such that 1 clauses of

are satisfied by some truth assignment.

3. If , then is (completely) satisfiable.

4. If , then no truth assignment satisfies more than 1 clauses of .


5. Given a truth assignmentwhich satisfiesmore than 1 clauses of , a truth assignment

which satisfies completely (or, alternatively, a witness showing ) can be constructed

in polynomial time.

Some of the properties above may not be immediately obvious from the construction given by

Arora, Lund, Motwani, Sudan, and Szegedy [7]. It is easy to verify that they provide a reduction

with properties (1), (3) and (4). Property (5) is obtained from the fact that all assignments which

satisfy most clauses are actually close (in terms of Hamming distance) to valid codewords from a

linear code, and the uniquely error-corrected codeword obtained from this “corrupted code-word”

will satisfy all the clauses of .

Property (2) requires a bit more care and we provide a brief sketch of how it may be ensured.

The idea is to revert back to the PCP model and redefine the proof verification game. Suppose that

the original game had the properties that for there exists a proof such that the verifier accepts

with probability 1, and otherwise, for , the verifier accepts with probability at most 1 2. We

now augment this game by adding to the proof a 0th bit which the prover uses as follows: if the bit

is set to 1, then the prover “chooses” to play the old game, else he is effectively “giving up” on the

game. The verifier in turn first looks at the 0th bit of the proof. If this is set, then she performs the

usual verification, else she tosses an unbiased coin and accepts if and only if it turns up heads. It is

clear that for there exists a proof on which the verifier always accepts. Also, for no

proof can cause the verifier to accept with probability greater than 1 2. Finally, by setting the 0th

bit to 0, the prover can create a proof which the verifier accepts with probability exactly 1 2. This

proof system can now be transformed into a 3-CNF formula of the desired form.

2.4.3 Reducing Maximization Problems to MAX 3-SAT

We have already established that, without loss of generality, we only need to worry about maxi-

mization problems Π APX-PB. Consider such a problem Π, and let be a polynomial-time

algorithm which delivers a -approximation for Π, where is some constant. Given any instance

of Π, let be the bound on the optimum value for obtained by running on input .

Note that this may be a stronger bound than the a priori polynomial bound on the optimum value


for any instance of length . An important consequence is that OPT .

We generate a sequence of NP decision problems OPT for 1 . Given

an instance , we create formulas , for 1 , using the reduction from Theorem 6, where

th formula is obtained from the NP language .

Consider now the formula 1 that has the following features.

The number of satisfiable clauses of is exactly

1 OPT

where and are as guaranteed by Theorem 6.

Given an assignment which satisfies 1 clauses of , we can construct in

polynomial time a solution to of value at least . To see this, observe the following:

any assignment which so many clauses must satisfy more than 1 clauses in at least

of the formulas . Let be the largest index for which this happens; clearly, .

Furthermore, by property (5) of Theorem 6, we can now construct a truth assignment which

satisfies completely. This truth assignment can be used to obtain a solution such that

.

In order to complete the proof it remains to be shown that given any truth assignment with

error , i.e., which satisfies 1 clauses of , we can find a solution for with error

for some constant . We show that this is possible for 2 . The main

idea behind finding such a solution is to use the second property above to find a “good” solution to

using a “good” truth assignment for .

Suppose we are given a solution which satisfies 1 clauses. Since 1

1 and 1 OPT , we can use the second feature from above

to construct a solution 1 such that

11 1

1 OPT


1 1 OPT

Suppose 1 . Let 1 and 1 . Then it is readily

seen that

1OPT1

and than

02

On the other hand, if 1 , then the error in a solution 2 obtained by running

the -approximation algorithm for Π is given by

12

Therefore, choosing 2 we immediately obtain that the solution with larger value,

among 1 and 2, has error at most . Thus, this reduction is indeed an -reduction.

2.5 Generic Reductions and an Approximation Hierarchy

In this section we describe a generic technique for turning a hardness result into an approximation

preserving reduction.

We start by listing the kind of constraints imposed on the hardness reduction, the approximation

class and the optimization problem. We will observe at the end that these restrictions are obeyed

by all known hardness results and the corresponding approximation classes.

Definition 21 (Additive Problems) An NPO problem Π is said to be additive if there exists an

operator and a polynomial time computable function such that maps a pair of instances

1 and 2 to an instance 1 2 such that OPT 1 2 OPT 1 OPT 2 , and maps

a solution to 1 2 to a pair of solutions 1 and 2 to 1 and 2, respectively, such that

1 2 1 1 2 2 .

Definition 22 (Downward Closed Family) A family of functions : is said


to be downward closed if for all and for all constants (and in particular for all integers

1), implies that . A function is said to be hard for the family if

for all , there exists a constant such that ; the function is said to be

complete for if is hard for and .

Definition 23 ( -APX) For a downward closed family , the class -APX consists of all polyno-

mially bounded optimization problems approximable to within a ratio of for some function

.

Definition 24 (Canonical Hardness) AnNPmaximization problemΠ is said to be canonically hard

for the class -APX if there exists a transformation mapping instances of 3-SAT to instances

of Π, constants 0 and , and a gap function which is hard for the family , such that given an

instance of 3-SAT on 0 variables and , is an instance of Π with the

following properties.

If 3-SAT, then OPT .

If 3-SAT, then OPT .

Given a solution to with , a truth assignment satisfying can be

found in polynomial time.

In the above definition, the transformation from 3-SAT to Π is somewhat special in that one

can specify the size/optimum of the reduced problem and can produce a mapped instance of the

desired size. This additional property is easily obtained for additive problems, by using sufficient

number of additions till the optimum of the reduced problem is close to the target optimum, and

then adding a problem of known optimum value to the reduced problem.

Canonical hardness for NP minimization problems is analogously defined: OPT when

the formula is satisfiable and OPT , otherwise. Given any solution with value less

than , one can construct a satisfying assignment in polynomial time.

2.5.1 The Reduction


Theorem 7 If is a downward closed family of functions, and an additive NPO problem Ω is

canonically hard for the class -APX-PB, then all problems in -APX-PB -reduce to Ω.

Proof: Let Π be a polynomially bounded optimization problem in -APX, approximable to

within by an algorithm , and let Ω be a problem shown to be hard to within a factor

where is hard for . Let and denote the objective functions of Π and Ω, respectively.

We start with the special case where both Π and Ω are maximization problems. We describe the

functions , and the constant as required for an -reduction.

Let Π be an instance of size ; pick so that is . To describe our reduction,

we need to specify the functions and . The function is defined as follows. Let .

For each 1 , let denote the NP-language OPT . Now for each ,

we create an instance Ω of size such that if then OPT is , and it is

otherwise. We define .

We now construct the function . Given an instance Π and a solution to , we

compute a solution to in the following manner. We first use to find a solution 1. We also

compute a second solution 2 to as follows. Let be the largest index such that the solution

projects down to a solution to the instance such that . This in turn

implies we can find a solution 2 to witness 2 . Our solution is the one among 1 and

2 that yields the larger objective function value.

We now show that the reduction is an -reduction with 1 1 .

Let OPT . Observe that

OPT

Consider the following two cases:

Case 1 [ ]: In this case, . Since is a solution to of error at most 1 ,

it suffices to argue that the error of as a solution to is at least 1 . We start with the

following upper bound on .


11

Thus the approximation factor achieved by is given by

1 11

1 11

1

So in this case 1 (and hence ) is a solution to with an error of at most , if is a solution to

with an error of .

Case 2 [ ]: Let . Note that 1 and that the error of as a solution to is

. We bound the value of the solution to as

and its error as

1

11 1

1

(The final inequality follows from the fact that 1 1 1 1 .)

Thus in this case also we find that (by virtue of 2) is a solution to of error at most if

is a solution to of of error .

We now consider the more general cases where Π and Ω are not both maximization problems.


For the case where both are minimization problems, the above transformation works with one minor

change. When creating , the NP language consists of instances such that there exists with

.

For the case where Π is a minimization problem and Ω is a maximization problem, we first

-reduce Π to a maximization problem Π and then proceed as before. The reduction proceeds

as follows. Since Π is a polynomially bounded optimization problem, we can compute an upper

bound on the value of any solution to an instance . Let be such a bound for an instance .

The objective function of Π on the instance is defined as 2 2 . To begin

with, it is easy to verify that Π -APX implies Π -APX.

Let be a solution to instance of Π of error . We will show that as a solution to instance

of Π has an error of at least 2. Assume, without loss of generality, that 0. Then

OPT OPT 1

Multiplying by 2 2 OPT , we get

2 2 2 2

OPT2 2

2

This implies that

2 2

OPT2 2

112

2 2

OPT2 2

12

2 2

Upon rearranging,

11 2

2 2

OPT1 1

1 22 2

OPT

Thus the reduction from Π to Π is an -reduction.

Finally, the last remaining case, i.e., Π being a maximization problem and Ω being a minimiza-

tion problem, is dealt with similarly: we transform Π into a minimization problem Π .


Remark 6 This theoremappears tomerge two different notions of the relative ease of approximation

of optimization problems. One such notion would consider a problem Π1 easier than Π2 if there

exists an approximation preserving reduction from Π1 to Π2. A different notion would regard Π1

to be easier than Π2 if one seems to have a better factor of approximation than the other. The

above statement essentially states that these two comparisons are indeed the same. For instance,

the MAX CLIQUE problem and the CHROMATIC NUMBER problem, which are both in poly-APX,

are inter-reducible to each other. The above observation motivates the search for other interesting

function classes , for which the class -APX may contain interesting optimization problems.

2.5.2 Applications

Theorem 7 can be used to obtain structural characterizations of the classes poly-APX and log-APX.

We need to introduce two other syntactic optimization classes.

Definition 25 (RMAX(k) [85]) RMAX is the class of NPO problems expressible as finding a

structure which maximizes the objective function

if Φ

0 otherwise

where is a single predicate and Φ is a quantifier-free CNF formula in which occurs at

most times in each clause and all its occurrences are negative.

The results of Panconesi and Ranjan [85] can be adapted to show that MAX CLIQUE is complete

under -reductions for the class RMAX(2).

Definition 26 (MIN F Π2 [73]) MIN F Π2 is the class of NPO problems expressible as

finding a structure which minimizes the objective function

: if Φ

0 otherwise


where is a single predicate, Φ is a quantifier-free CNF formula in which occurs at

most times in each clause and all its occurrences are positive.

The results of Kolaitis and Thakur [73] can be adapted to show that SET COVER is complete

under -reductions for the class MIN F Π2 1 .

The following may now be concluded from Theorem 7.

Theorem 8

a) RMAX(2) poly-APX.

b) If SETCOVER is canonically hard to approximate towithin a factor ofΩ log , then log-APX

MIN F Π2 1 .

We briefly sketch the proof of this theorem. The hardness reduction for MAX SAT and CLIQUE

are canonical [7, 37]. The classes APX-PB, poly-APX, log-APX are expressible as classes -APX

for downward closed function families. The problems MAX SAT, MAX CLIQUE and SET COVER

are additive. Thus, we can now apply Theorem 7.

Remark 7 We would like to point out that almost all known instances of hardness results seem to

be shown for problems which are additive. In particular, this is true for all MAX SNP problems,

MAX CLIQUE, CHROMATIC NUMBER, and SET COVER. Two cases where a hardness result

does not seem to directly apply to an additive problem is that of LONGEST PATH [64] and BIN

PACKING. In the former case, the closely related LONGEST - PATH problem is easily seen to

be additive and the hardness result essentially stems from this problem. As for the case of BIN

PACKING, which does not admit a PTAS, the hardness result is not of a multiplicative nature and in

fact this problem can be approximated to within arbitrarily small factors, provided a small additive

error term is allowed. This yields a reason why this problem will not be additive. Lastly, the most

interesting optimization problems which do not seem to be additive are problems related to GRAPH

BISECTION or PARTITION, and these also happen to be notable instances where no hardness of

approximation results have been achieved!


2.6 Local Search and MAX SNP

In this section we present a formal definition of the paradigm of non-oblivious local search. The

idea of non-oblivious local search has been implicitly present in some well-known techniques such

as the interior-point methods. We will formalize this idea in context of MAX SNP and illustrate its

application to MAX SNP problems. Given a MAX SNP problem Π, recall that the goal is to find a

structure which maximizes the objective function: Φ . In the subsequent

discussion, we view as a -dimensional boolean vector.

2.6.1 Classical Local Search

We start by reviewing the standard mechanism for constructing a local search algorithm. A

-local algorithm for Π is based on a distance function 1 2 which is the Hamming

distance between two -dimensional vectors. The -neighborhood of a structure is given by

, where is the universe. A structure is called -optimal if

, we have . The algorithm computes a -optimum by performing

a series of greedy improvements to an initial structure 0, where each iteration moves from the

current structure to some 1 of better value (if any). For constant , a -local

search algorithm for a polynomially-bounded NPO problem runs in polynomial time because:

each local change is polynomially computable, and

the number of iterations is polynomially bounded since the value of the objective function

improves monotonically by an integral amount with each iteration, and the optimum is

polynomially-bounded.

2.6.2 Non-Oblivious Local Search

A non-oblivious local search algorithm is based on a 3-tuple 0 , where 0 is the initial

solution structure which must be independent of the input, is a real-valued function referred

to as theweight function, and is a real-valued distance functionwhich returns the distance between

two structures in some appropriately chosen metric. The weight function should be such that the


number of distinct values taken by is polynomiallybounded in the input size. Moreover, the

distance function should be such that given a structure and a fixed , can be computed

in time polynomial in . Then, as in classical local search, for constant , a non-oblivious -local

algorithm terminates in time polynomial in the input size.

The classical local search paradigm, which we call oblivious local search, makes the natural

choice for the function , and the distance function , i.e., it chooses them to be

and the Hamming distance. However, as we show later, this choice does not always yield a good

approximation ratio. We now formalize our notion of this more general type of local search.

Definition 27 (Non-Oblivious Local Search Algorithm) A non-oblivious local search algorithm

is a -local search algorithm whose weight function is defined to be

1Φ

where is a constant, Φ ’s are quantifier-free first-order formulas, and the profits are real

constants. The distance function is an arbitrary polynomial-time computable function.

A non-oblivious local search can be implemented in polynomial time in much the same way

as the oblivious local search. Note that the we are only considering polynomially-bounded weight

functions and the profits are fixed independent of the input size. In general, the non-oblivious

weight functions do not direct the search in the direction of the actual objective function. In

fact, as we will see, this is exactly the reason why they are more powerful and allow for better

approximations.

We now define two classes of NPO problems.

Definition 28 (Oblivious GLO) The class of problems OBLIVIOUS GLO consists of allNPO prob-

lems which can be approximated within constant factors by an oblivious -local search algorithm

for some constant .

Definition 29 (Non-Oblivious GLO) The class of problems NON-OBLIVIOUS GLO consists of all

NPO problemswhich can be approximatedwithin constant factors by a non-oblivious -local search

algorithm for some constant .


Remark 8 It would be perfectly reasonable to allow weight functions that are non-linear, but we

stay with the above definition for the purposes of our study. Allowing only a constant number of

predicates in the weight functions enables us to prevent the encoding of arbitrarily complicated

approximation algorithms. The structure is a -dimensional vector, and so a natural metric for

the distance function is the Hamming distance. In fact, classical local search is indeed based on

the Hamming metric and this is useful in proving negative results for the paradigm. In contrast,

the definition of non-oblivious local search allows for other distance functions, but we will use

only the Hamming metric in proving positive results in the remainder of this chapter. However,

we have found that it is sometimes useful to modify this, for example by modifying the Hamming

distance so that the complement of a vector is considered to be at distance 1 from it. Finally, it is

sometimes convenient to assume that the local search makes the best possible move in the bounded

neighborhood, rather than an arbitrary move which improves the weight function. We believe that

this does not increase the power of non-oblivious local search.

2.7 The Power of Non-Oblivious Local Search

We will show that there exists a choice of a non-oblivious weight function for MAX k-SAT such

that any assignment which is 1-optimal with respect to this weight function, yields a performance

ratio of 2 2 1 with respect to the optimal. But first, we obtain tight bounds on the performance

of oblivious local search for MAX 2-SAT, establishing that its performance is significantly weaker

than the best-known result even when allowed to search exponentially large neighborhoods. We use

the following notation: for any fixed truth assignment , is the set of clauses in which exactly

literals are true; and, for a set of clauses , denotes the total weight of the clauses in .

2.7.1 Oblivious Local Search forMAX 2-SAT

We show a strong separation in the performance of oblivious and non-oblivious local search for

MAX 2-SAT. Suppose we use a -local strategy with the weight function being the total weight

of the clauses satisfied by the assignment, i.e., 1 2 . The following theorem

shows that for any , an oblivious -local strategy cannot deliver a performance ratio better


than 3 2. This is rather surprising given that we are willing to allow near-exponential time for the

oblivious algorithm.

Theorem 9 The asymptotic performance ratio of an oblivious -local search algorithm for the

MAX 2-SAT problem is 3 2 for any positive integral . This ratio is still bounded by 5 4

when may take any value less than 2.

Proof: We first show the existence of an input instance for MAX 2-SAT which may elicit a

relatively poor performance ratio for any -local algorithm provided . In our construction

of such an input instance, we assume that 2 1. The input instance comprises of a disjoint

union of four sets of clauses, say Γ1 Γ2 Γ3 and Γ4, defined as below:

Γ11

Γ21

Γ30

2 1

Γ42 2

Clearly, Γ1 Γ2 2 , and Γ3 Γ4 2 1 . Without loss of generality,

assume that the current input assignment is 1 1 1 . This satisfies all clauses in Γ1 and

Γ2. But none of the clauses in Γ3 and Γ4 are satisfied. If we flip the assignment of values to any

variables, it would unsatisfy precisely clauses in Γ1 Γ2. This is the number of

clauses in Γ1 Γ2 where a flipped variable occurs with an unflipped variable.

On the other hand, flipping the assigned values of any variables can satisfy at most

clauses in Γ3 Γ4 as we next show.

Let Π denote the set of clauses on variables given by 0 2 1 2 2

where 2 1 . We claim the following.


Lemma 1 Any assignment of values to the variables such that at most variables have been

assigned value false, can satisfy at most clauses in Π .

Proof: We prove by simultaneous induction on and that the statement is true for any instance

Π where and are non-negative integers such that 2 1 . The base case includes

1 and 2 and is trivially verified to be true for the only allowable value of , namely

0. We now assume that the statement is true for any instance Π such that and

2 1 . Consider now the instance Π . The statement is trivially true for 0. Now

consider any 0 such that 2 1 . Let 1 2 be any choice of variables

such that for . Again the assertion is trivially true if 0 or 1. We assume that

2 from now on. If we delete all clauses containing the variables 1 and 2 from Π , we

get the instance Π 2 1 . We now consider three cases.

Case 1 [ 1 3]: In this case, we are reduced to the problem of finding an upper bound on the

maximum number of clauses satisfied by setting any variables to false in Π 2 1 . If

1, we may use the inductive hypothesis to conclude that no more than 2

clauses will be satisfied. Thus the assertion holds in this case. However, we may not directly use the

inductive hypothesis if . But in this case we observe that since by the inductive hypothesis,

setting any 1 variables in Π 2 1 to false, satisfies at most 2 1 1

clauses, assigning the value false to any set of variables, can satisfy at most

2 1 11

12 1 1 2

clauses. Hence the assertion holds in this case also.

Case 2 [ 1 2]: In this case, 1 satisfies one clause and the remaining 1 variables satisfy at

most 2 1 1 clauses by the inductive hypothesis on Π 2 1 . Adding up

the two terms, we see that the assertion holds.

Case 3 [ 1 1]: We analyze this case based on whether 2 2 or 2 3. If 2 2, then 1

and 2, together satisfy precisely 1 clauses and the remaining 2 variables, satisfy at most


2 2 2 clauses using the inductive hypothesis. Thus the assertion still holds.

Otherwise, 1 satisfies precisely 1 clauses and the remaining 1 variables satisfy no more

than 1 1 1 clauses using the inductive hypothesis. Summing up the two terms,

we get as the upper bound on the total number of clauses satisfied. Thus the assertion

holds in this case also.

To see that this bound is tight, simply consider the situation when the variables set to

false are 1 3 2 1, for any . The total number of clauses satisfied is given by

1 2 1 .

Assuming that each clause has the same weight, Lemma 1 allows us to conclude that a -local

algorithm cannot increase the total weight of satisfied clauses with this starting assignment. An

optimal assignment on the other hand can satisfy all the clauses by choosing the vector

0 0 0 . Thus the performance ratio of a -local algorithm, say , is bounded as

Γ1 Γ2 Γ3 Γ4Γ1 Γ2

3 2 12 2

For any , this ratio asymptotically converges to 3 2. We next show that this bound is

tight since a 1-local algorithm achieves it. However, before we do so, we make another intriguing

observation, namely, for any 2, the ratio is bounded by 5 4.

Now to see that a 1-local algorithm ensures a performance ratio of 3 2, consider any 1-optimal

assignment and let denote the set of clauses containing the variable such that no literal in

any clause of is satisfied by . Similarly, let denote the set of clauses containing the variable

such that precisely one literal is satisfied in any clause in and furthermore, it is precisely

the literal containing the variable . If we complement the value assigned to the variable , it is

exactly the set of clauses in which becomes satisfied and the set of clauses in which is no

longer satisfied. Since is 1-optimal, it must be the case that . If we sum up

this inequality over all the variables, then we get the inequality 1 1 . We

observe that 1 2 0 and 1 1 because each clause in 0 gets

counted twice while each clause in 1 gets counted exactly once. Thus the fractional weight of the


number of clauses not satisfied by a 1-local assignment is bounded as

0

0 1 2

03 0 2

03 0

13

Hence the performance ratio achieved by a 1-local algorithm is bounded from above by 3 2.

Combining this with the upper bound derived earlier, we conclude that 1 3 2. This concludes

the proof of the theorem.

2.7.2 Non-Oblivious Local Search for MAX 2-SAT

We now illustrate the power of non-oblivious local search by showing that it achieves a performance

ratio of 4 3 for MAX 2-SAT, using 1-local search with a simple non-oblivious weight function.

Theorem 10 Non-oblivious 1-local search achieves a performance ratio of 4 3 for MAX 2-SAT.

Proof: We use the non-oblivious weight function

32 1 2 2

Consider any assignment which is 1-optimal with respect to this weight function. Without loss

of generality, we assume that the variables have been renamed such that each unnegated literal

gets assigned the value true. Let and respectively denote the total weight of clauses in

containing the literals and , respectively. Since is a 1-optimal assignment, each variable

must satisfy the following equation.

12 2

32 1

12 1

32 0 0

Summing this inequality over all the variables, and using

11

11 1

12 2 2


10 2 0

we obtain the following inequality:

2 1 3 0

This immediately implies that the total weight of the unsatisfied clauses at this local optimum is no

more than 1 4 times the total weight of all the clauses. Thus, this algorithm ensures a performance

ratio of 4 3.

Remark 9 The same result can be achieved by using the oblivious weight function, and instead

modifying the distance function so that it corresponds to distances in a hypercube augmented by

edges between nodes whose addresses are complement of each other.

2.7.3 Generalization toMAX k-SAT

We can also design a non-oblivious weight function for MAX k-SAT such that a 1-local strategy

ensures a performance ratio of 2 2 1 . The weight function will be of the form

0 where the coefficients ’s will be specified later.

Theorem 11 Non-oblivious 1-local search achieves a performance ratio of2 2 1 forMAX k-SAT.

Proof: Again, without loss of generality, we will assume that the variables have been renamed

so that each unnegated literal is assigned true under the current truth assignment. Thus the set is

the set of clauses with unnegated literals.

Let ∆ 1 and let denote the change in the current weight when we flip the value

of , that is, set it to 0. It is easy to verify the following equation:

∆2

∆ 1 ∆ 1 1 ∆1 0 (2.1)

Thus when the algorithm terminates, we know that 0, for 1 . Summing over

all values of , and using the fact 1 and 1 we get the


following inequality.

∆2

1∆ ∆ 1 ∆1 0 (2.2)

We now determine the values of ∆ ’s such that the coefficient of each term on the left hand side

is unity. It can be verified that

∆1

1 1 0

achieves this goal. Thus the coefficient of 0 on the right hand side of equation (2.2) is 2 1.

Clearly, the weight of the clauses not satisfied is bounded by 1 2 times the total weight of all the

clauses. It is worthwhile to note that this is regardless of the value chosen for the coefficient 0.

2.8 Local Search for CSP and MAX SNP

We now introduce a class of constraint satisfaction problems such that the problems in MAX SNP

are exactly equivalent to the problems in this class. Furthermore, every problem in this class can

be approximated to within a constant factor by a non-oblivious local search algorithm.

2.8.1 Constraint Satisfaction Problems

The connection between the syntactic description of optimization problems and their approxima-

bility through non-oblivious local search is made via a problem called MAX k-CSP which captures

all the problems in MAX SNP as a special case.

Definition 30 (k-ary Constraint) Let 1 be a set of boolean variables. A -ary

constraint on is ; , where is a size subset of , and : is a

-ary boolean predicate.

Definition 31 (MAX k-CSP) Given a collection 1, , of weighted -ary constraints over

the variables 1 , the MAX k-CSP problem is to find a truth assignment satisfying a

maximum weight sub-collection of the constraints.


The following theorem shows that MAX k-CSP problem is a “universal” MAX SNP problem,

in that it contains as special cases all problems in MAX SNP.

Theorem 12

a) For fixed , MAX k-CSP MAX SNP.

b) Let Π MAX SNP. Then, for some constant , Π is a MAX k-CSP problem. Moreover, the

-CSP instance corresponding to any instance of this problem can be computed in polynomial

time.

Proof: The proof of part (b) is implicit in Theorem 1 in [87], and so we concentrate on proving

part (a). Our goal is to obtain a representation of the -CSP problem in the MAX SNP syntax:

max Φ

The input structure is MAX; ARG EVAL , where 1 , MAX

contains the integers 1 max , the predicate ARG encodes the sets , and the predicate

EVAL encodes the predicates , as described below.

ARG is 3-ary predicate which is true if and only if the th argument of is the

variable , for 1 , 1 , and 1 .

EVAL 1 is a 1 -ary predicate which is true if and only if 1

evaluates to true, for 1 and all .

The structure is defined as ; TRUE , where TRUE is a unary predicate which denotes an

assignment of truth values to the variables in . The vector has 1 components which will be

called 1, , and , for convenience. The intention is that the ’s refer to the arguments of the

th constraint.

All that remains is to specify the quantifier-free formula Φ. The basic idea is that Φ

should evaluate to true if and only if the following two conditions are satisfied:

the arguments of the constraint are given by the variables 1, , , in that order; and,


the values given to these variables under the truth assignment specified by are such that the

constraint is satisfied.

The formula Φ is given by the following expression, with the two sub-formulas ensuring these two

conditions.

1ARG

1

EVAL 11

TRUE

It is easy to see that the first sub-formula has the desired effect of checking that the ’s correspond

to the arguments of . The second sub-formula considers all possible truth assignment to these

variables, and checks that the particular set of values assigned by the structure will make

evaluate to true.

For a fixed structure , there is exactly one choice of per constraint that could makeΦ evaluate

to true, and this is happens if and only if that constraint is satisfied. Thus, the value of the solution

given by any particular truth assignment structure is exactly the number of constraints that are

satisfied. This shows that the MAX SNP problem always has the same value as intended in the

-CSP problem.

Finally, there are still a few things which need to be checked to ensure that this is a valid

MAX SNP formulation. Notice that all the predicates are of bounded arity and the structures

consist of a bounded number of such predicates, i.e., independent of the input size which is given

by MAX. Further, although the length of the formula is exponential in , it is independent of the

input.

2.8.2 Non-Oblivious Local Search for MAX k-CSP

A suitable generalization of the non-oblivious local search algorithm for MAX k-SAT yields the

following result.

Theorem 13 A non-oblivious 1-local search algorithm has performance ratio 2 for MAX k-CSP.

Proof: We use an approach similar to the one used in the previous section to design a non-

oblivious weight function for the weighted version of the MAX k-CSP problem such that a 1-local


algorithm yields 2 performance ratio to this problem.

We consider only the constraints with at least one satisfying assignment. Each such constraint

can be replaced by a monomial which is the conjunction of some literals such that when the

monomial evaluates to true the corresponding literal assignment represents a satisfying assignment

for the constraint. Furthermore, each such monomial has precisely one satisfying assignment. We

assign to each monomial the weight of the constraint it represents. Thus any assignment of variables

which satisfies monomials of total weight 0, also satisfies constraints in the original problem of

total weight 0.

Let denote the monomials with true literals, and assume that the weight function is of

the form 1 . Thus, assuming that the variables have been renamed so that the current

assignment gives value true to each variable, we know that for any variable , is given by

equation (2.1). As before, using the fact that for any 1-optimal assignment, 0 for 1 ,

and summing over all values of , we can write the following inequality.

∆1 0

1

2∆ 1 ∆ ∆ (2.3)

We now determine the values of ∆ ’s such that the coefficient of each term on the left hand side is

unity. It can be verified that

∆1 1

0

achieves this goal. Thus the coefficient of on the right hand side of equation (2.1) is 2 1.

Clearly, the total weight of clauses satisfied is at least 1 2 times the total weight of all the clauses

with at least one satisfiable assignment.

We conclude the following theorem.

Theorem 14 Every optimization problem Π MAX SNP can be approximated to within some

constant factor by a (uniform) non-oblivious 1-local search algorithm, i.e.,

MAX SNP NON-OBLIVIOUS GLO

For a problem expressible as -CSP, the performance ratio is at most 2 .


2.9 Non-Oblivious versus Oblivious GLO

In this section, we show that there exist problems for which no constant factor approximation can

be obtained by any -local search algorithm with oblivious weight function, even when we allow

to grow with the input size. However, a simple 1-local search algorithm using an appropriate

non-oblivious weight function can ensure a constant performance ratio.

2.9.1 MAX 2-CSP

The first problem is an instance of MAX 2-CSP where we are given a collection of monomials such

that each monomial is an “and” of precisely two literals. The objective is to find an assignment to

maximize the number of monomials satisfied.

We show an instance of this problem such that for every , there exists an instance one

of whose local optima has value that is a vanishingly small fraction of the global optimum.

The input instance consists of a disjoint union of two sets of monomials, say Γ1 and Γ2, defined

as below:

Γ11

Γ21

Clearly, Γ1 2 , and Γ21

2 . Consider the truth assignment 1 1 1 .

It satisfies all monomials in Γ2 but none of the monomials in Γ1. We claim that this assignment is

-optimal with respect to the oblivious weight function. To see this, observe that complementing the

value of any variables will unsatisfy at least 2 monomials in Γ2 for any . On the

other hand, this will satisfy precisely 2 monomials in Γ1. For any , we have 2 2 ,

and so is a -local optimum.

The optimal assignment on the other hand, namely OPT 0 0 0 , satisfies all mono-

mials in Γ1. Thus, for 2, the performance ratio achieved by any -local algorithm is no more

than 21

2 which asymptotically diverges to infinity for any . We have already

seen in Section 2.8 that a 1-local non-oblivious algorithm ensures a performance ratio of 4 for this


problem. Since this problem is in MAX SNP, we obtain the following theorem.

Theorem 15 There exist problems in MAX SNP such that for , no -local oblivious

algorithm can approximate them to within a constant performance ratio, i.e.,

MAX SNP OBLIVIOUS GLO

2.9.2 Vertex Cover

Ausiello and Protasi [10] have shown that VERTEX COVER does not belong to the class GLO

and, hence, there does not exist any constant such that an oblivious -local search algorithm

can compute a constant factor approximation. In fact, their example can be used to show that for

any , the performance ratio ensured by -local search asymptotically diverges to infinity.

However, we show that there exists a rather simple non-oblivious weight function which ensures a

factor 2 approximation via a 1-local search. In fact, the algorithm simply enforces the behavior of

the standard approximation algorithm which iteratively builds a vertex cover by simply including

both end-points of any currently uncovered edge.

We assume that the input graph is given as a structure where is the set of vertices

and encodes the edges of the graph. Our solution is represented by a 2-ary predicate

which is iteratively constructed so as to represent a maximal matching. Clearly, the end-points

of any maximal matching constitute a valid vertex cover and such a vertex cover can be at most

twice as large as any other vertex cover in the graph. Thus is an encoding of the vertex cover

computed by the algorithm.

The algorithm starts with initialized to the empty relation and at each iteration, at most one

new pair is included in it. The non-oblivious weight function used is as below:

3

Φ1 2Φ2 Φ3

where

Φ1


Φ2

Φ3

Let encode a valid matching in the graph . We make the following observations.

Any relation obtained from by either deleting an edge from it, or including an edge

which is incident on an edge of , or including a non-existent edge, has the property that

. Thus in a 1-local search from , we will never move to a relation

which does not encode a valid matching of .

On the other hand, if a relation corresponds to the encoding of a matching in which is

larger than the matching encoded by , then . Thus if does not

encode a maximal matching in , there always exist a relation in its 1-neighborhood of larger

weight than itself.

These two observations, combined with the fact that we start with a valid initial matching (the

empty matching), immediately allow us to conclude that any 1-optimal relation always encodes

a maximal matching in . We have established the following.

Theorem 16 A 1-local search algorithm using the above non-oblivious weight function achieves a

performance ratio of 2 for the VERTEX COVER problem.

Theorem 17 GLO is a strict subset of NON-OBLIVIOUS GLO

As an aside, it can be seen that this algorithm has the same performance starting with an arbitrary

initial solution. This is because for any relation not encoding a matching of the input graph,

deleting one of the violating members strictly increases .

2.10 The Traveling Salesman Problem

The TSP(1,2) problem is the traveling salesman problem restricted to complete graphs where all

edge weights are either 1 or 2; clearly, this satisfies the triangle inequality. Papadimitriou and


Yannakakis [88] showed that this problem is hard for MAX SNP. The natural weight function for

TSP(1,2), that is, the weight of the tour, can be used to show that a 4-local algorithm yields a 3 2

performance ratio. The algorithm starts with an arbitrary tour and in each iteration, it checks if there

exist two disjoint edges and on the tour such that deleting them and replacing them with

the edges and yields a tour of lesser cost.

Theorem 18 A 4-local search algorithm using the oblivious weight function achieves a 3 2 per-

formance ratio for TSP(1,2).

Proof: Let be a 4-optimal solution and let be a permutation such that the vertices in

occur in the order 1 2 . Consider any optimal solution . With each unit cost edge

in , we associate a unit cost edge in as follows. Let where . If 1

then . Otherwise, consider the edges 1 1 and 2 1 on . We claim

either 1 or 2 must be of unit cost. Suppose not, then the tour which is obtained by simply

deleting both 1 and 2 and inserting the edges and 1 1 has cost at least one less

than . But is 4-optimal and thus this is a contradiction.

Let denotes the set of unit cost edges in and let be the set of unit cost edges in

which form the image of under the above mapping. Since an edge 1 in can

only be the image of unit cost edges incident on in and since is a tour, there are at most

two edges in which map to . Thus 2 and hence

22 2 2

23

In fact, the above bound can be shown to be tight.

2.11 Maximum Independent Sets in Bounded Degree Graphs

The input instance to the maximum independent set problem in bounded degree graphs, denoted

MIS-B, is a graph such that the degree of any vertex in is bounded by a constant ∆. We present

an algorithm with performance ratio 8∆2 4∆ 1 2∆ 1 2 for this problem when ∆ 10.


Our algorithm uses two local search algorithms such that the larger of the two independent sets

computed by these algorithms, gives us the above claimed performance ratio. We refer to these two

algorithms as 1 and 2.

In our framework, the algorithm 1 can be characterized as a 3-local algorithm with the weight

function simply being 3 . Thus if we start with initialized to empty set, it is

easy to see that at each iteration, will correspond to an independent set in . A convenient way

of looking at this algorithm is as follows. We define an swap to be the process of deleting

vertices from and including vertices from the set to the set . In each iteration, the

algorithm 1 performs either a 0 swap where 1 3, or a 1 2 swap. A 0 swap

however, can be interpreted as applications of 0 1 swaps. Thus the algorithm may be viewed

as executing a 0 1 swap or a 1 2 swap at each iteration. The algorithm terminates when

neither of these two operations is applicable.

Let denote the 3-optimal independent set produced by the algorithm 1. Furthermore, let

be any optimal independent set and let . We make the following useful observations.

Since for no vertex in , a 0 1 swap can be performed, it implies that each vertex in

must have at least one incoming edge to .

Similarly, since no 1 2 swaps can be performed, it implies that at most vertices

in can have precisely one edge coming into . Thus

vertices in must have at least two edges entering the set .

A rather straightforward consequence of these two observations is the following lemma.

Lemma 2 The algorithm 1 has performance ratio ∆ 1 2 for MIS-B.

Proof: The above two observations imply that the minimum number of edges entering from

the vertices in is 2 . On the other hand, the maximum number of edges

coming out of the vertices in to the vertices in is bounded by ∆. Thus we must

have

∆ 2


Rearranging, we get2

∆ 1∆ 1∆ 1

which yields the desired result.

This nearly matches the approximation ratio of ∆ 2 due to Hochbaum [53]. It should be noted

that the above result holds for a broader class of graphs, viz., -claw free graphs. A graph is called

-claw free if there does not exist an independent set of size or larger such that all the vertices in

the independent set are adjacent to the same vertex. Lemma 2 applies to ∆ 1 -claw free graphs.

Our next objective is to further improve this ratio by using the algorithm 1 in combination

with the algorithm 2. The following lemma uses a slightly different counting argument to give an

alternative bound on the approximation ratio of the algorithm 1 when there is a constraint on the

size of the optimal solution.

Lemma 3 For any real number ∆, the algorithm 1 has performance ratio ∆ 2 for

MIS-B when the optimal value itself is no more than ∆ ∆ 4 .

Proof: As noted earlier, each vertex in must have at least one edge coming into the set

and at least vertices in must have at least two edges coming into . Therefore, the

following inequality must be satisfied:

∆

Thus ∆ 2 . Finally, observe that

∆ 22

∆

whenever ∆ ∆ 4 .

The above lemma shows that the algorithm 1 yields a better approximation ratio when the size

of the optimal independent set is relatively small.

The algorithm 2 is simply the classical greedy algorithm. This algorithm can be conveniently

included in our framework if we use directed local search. If we let denote the set of neighbors


of the vertices in , then the weight function is simply ∆ 1

∆ 1 . It is not difficult to see that starting with an empty independent set, a 1-local

algorithm with directed search on above weight function simply simulates a greedy algorithm. The

greedy algorithm exploits the situation when the optimal independent set is relatively large in size.

It does so by using the fact that the existence of a large independent set in ensures a large subset

of vertices in with relatively small average degree. The following two lemmas characterize the

performance of the greedy algorithm.

Lemma 4 Suppose there exists an independent set such that the average degree of vertices

in is bounded by . Then for any 1, the greedy algorithm produces an independent set of

size at least 1 .

Proof: The greedy algorithm iteratively chooses a vertex of smallest degree in the remaining

graph and then deletes this vertex and all its neighbors from the graph. We examine the behavior

of the greedy by considering two types of iterations. First consider the iterations in which it picks

a vertex outside . Suppose in the th such iteration, it picks a vertex in with exactly

neighbors in the set in the remaining graph. Since each one of these vertices must also have

at least edges incident on them, we loose at least 2 edges incident on . Suppose only such

iterations occur and let 1 . We observe that 12 . Secondly, we consider

the iterations when the greedy selects a vertex in . Then we do not loose any other vertices in

because is an independent set. Thus the total size of the independent set constructed by the

greedy algorithm is at least where .

By the Cauchy-Schwartz inequality, 12 2 . Therefore, we have 1

2 . Rearranging, we obtain that

2

1

2

1 1

2

12

1

Thus

1

2

12

1

But 2 1 for 1, and the result follows.


Lemma 5 For ∆ 10 and any non-negative real number 3∆ 8∆2 4∆ 1 1, the

algorithm 2 has performance ratio ∆ 2 for MIS-B when the optimal value itself is at least

∆ ∆ 4 .

Proof: Observe that the average degree of vertices in is bounded by ∆ and

thus using the fact that ∆ ∆ 4 , we know that the algorithm 2 computes an

independent set of size at least 1 where 4∆ 2∆ ∆ , and 1 for 0.

Hence it is sufficient to determine the range of values can take such that the following inequality

is satisfied:

12

∆

Substituting the bound on the value of and rearranging the terms of the equation, yields the

following quadratic equation :

2 6∆ 2 ∆2 10∆ 0

Since must be strictly bounded by ∆, the above quadratic equation is satisfied for any choice

of 3∆ 8∆2 4∆ 1 1 if ∆ 10.

Combining the results of Lemmas 3 and 5 and choosing the largest allowable value for , we

get the following result.

Theorem 19 An approximation algorithm which simply outputs the larger of the two independent

sets computed by the algorithms 1 and 2, has performance ratio 8∆2 4∆ 1 2∆ 1 2

for MIS-B.

The performance ratio claimed above is essentially ∆ 2 414. This improves upon the long-

standing approximation ratio of ∆ 2 due to Hochbaum [53], when ∆ 10. However, more

recently, there has been a flurry of new results for this problem. Berman and Furer [18] have given

an algorithm with performance ratio ∆ 3 5 when ∆ is even, and ∆ 3 25 5 for odd

∆, where 0 is a fixed constant. Halldorsson and Radhakrishnan [50] have shown that algorithm

1 when run on -clique free graphs, yields an independent set of size at least 2 ∆ . They


combine this algorithm with a clique-removal based scheme to achieve a performance ratio of

∆ 6 1 1 .

In conclusion, note that Khanna, Motwani and Vishwanathan [69] have recently shown that a

semi-definite programming technique can be used to obtain a ∆ log log∆ log∆ -approximation

algorithm for this problem.

Chapter 3

Towards a Structural Characterization

of PTAS

61

CHAPTER 3. TOWARDS A STRUCTURAL CHARACTERIZATION OF PTAS 62

3.1 Introduction

In the preceding chapter, we saw that we can use syntactic classes to provide a structural basis for

many natural approximation classes such as APX, log-APX and poly-APX. The starting point of

our study was the availability of syntactic classes which were contained in these approximation

classes and contained problems which were hard to approximate beyond the factors allowed for

in the definition of the corresponding classes. In this context, it is interesting to attempt a similar

characterization of other approximation classes as the closure of a syntactic class. In this chapter,

we focus on the class PTAS. The starting point for our investigation is the remarkable paper by

Baker [13] which presents PTAS’s for a surprisingly large number of NP-hard graph optimization

problems when the input graph is planar. It is our contention that the defining characteristic of

PTAS problems is a “restriction on input structure” for some syntactic classes. Our work provides

evidence towards this claim, raising the possibility that this input restriction may well be a notion

related to planarity.

This may appear surprising at first considering that several well-known PTAS problems do not

exhibit a graph structure, leave alone a planar graph structure, e.g., knapsack and multiprocessor

scheduling. However, we will demonstrate that these problems are in fact special cases of generic

problems arising from a planar input-restriction of generalizations of classes such as MAX SNP.

It appears that most problems in PTAS fit directly into our framework. Moreover, our framework

trivially implies membership in PTAS for problems such as PLANAR MAX SAT that were not known

to be in PTAS earlier.

We identify a family of syntactic classes with input restrictions that subsume planar-restricted

versions of classes such as MAX SNP, RMAX(2), and MIN F Π2 1 . We then demonstrate that

these classes are contained in PTAS using a collection of techniques that considerably generalize

(and simplify) the idea used by Baker [13]. Our techniques are based on a separator theorem for

outerplanar graphs due to Bodlaender [21] that strengthens the planar separator theorem of Lipton

and Tarjan [79]. While almost every known problem in PTAS is easily seen to belong to our syntactic

classes, there are a few problems, namely bin packing, planar graph TSP and some optimization

This chapter is based on joint work with Rajeev Motwani [68].


problems on dense instances, that do not fit into our framework. We point out that, strictly speaking,

many of these problems do not belong to PTAS, and hence would not serve as counter-examples to

the possibility that some extension of our framework (via closure) is a characterization of PTAS.

Before describing our results in greater detail, we briefly review some earlier results concerning

the role of planarity in PTAS. Lipton and Tarjan [80] were the first to present a PTAS for an NP-

hard optimization problem on planar graphs, namely, maximum independent sets (MIS) in planar

graphs. Subsequently, Chiba, Nishizeki, and Saito [24] extended this to finding a maximum induced

subgraph satisfying a hereditary property determined by connected components.

The essential idea in results of this nature is that a recursive application of the planar separator

theorem allows one to delete vertices from a given planar graph , so as to create a graph

such that every component of has size log . Then, it is possible to compute optimal solutions

in each component in polynomial time, and the union of these optimal solutions is guaranteed to

be a solution to the original problem. Using the fact that any such property must have a solution

of size Ω , we can conclude that the resulting solution is a 1 approximation, for any fixed

0 and sufficiently large .

Hunt et al [54] applied Baker’s technique to satisfiability problems where each clause consists

of an application of a bounded-arity function from a given set of boolean functions. They also

showed that for certain MAX SNP-complete problems, such as bounded-degree MIS, Baker’s

results can be extended to -near-planar graphs, i.e., graphs embeddable in the plane with at most

edge crossings. They also show that Baker’s framework can be suitably adapted to yield parallel

approximation schemes.

Recently, Grigni, Koustsoupias, and Papadimitriou [48] gave a PTAS for TSP on unweighted

planar graphs. Their algorithm combines a new separator theorem with a careful decomposition to

create such an approximation scheme. Soon thereafter, Arora [3] showed a PTAS for the Euclidean

TSP problem.

The rest of this chapter is organized as follows. Section 3.2 gives an overview of the results

and discusses some applications and directions for extending our results. Section 3.3 reviews basic

concepts related to outerplanar graphs and tree decompositions. Finally, in Sections 3.4, 3.5, 3.6,

and 3.7, we sketch the PTAS results for our proposed classes.


3.2 Summary of Results and the Implications

We begin by defining a new family of syntactic subclasses of NPO that subsume the known

syntactically-defined optimization classes such as MAX SNP, MAX NP, RMAX and MIN F Π2

(see Chapters 1 and 2 for definitions). Let 1 be a set of boolean variables. A positive

literal is a variable , and a negative literal is a negated variable . We use the abbreviation FOF to

denote quantifier-free first-order formulas over these literals. We associate a weight with each FOF

and with each variable; depending on the problem at hand, we may refer to the weights as profits or

costs. The weight of a collection of FOF’s for a given truth assignment is the net weight of the

satisfied formulas, and the weight of is the net weight of the variables assigned TRUE.

Definition 32 (Positive and Negative Minterms)Aminterm is any conjunction of literals. Aminterm

is positive if it has only positive literals, and is negative if it has only negative literals.

The following are the new syntactic classes of optimization problems. The class MPSAT

captures constraint satisfaction problems where the objective is to compute solutions satisfying as

many constraints as possible. It contains the usual satisfiability problems such as MAX 3-SAT. The

other two classes, TMAX and TMIN, capture problems where a feasible solution must satisfy given

constraints, and the objective is to find extremal feasible solutions. For instance, the MIS problem

requires a solution containing an extremal number of vertices that satisfy constraints ruling out the

inclusion of both end-points of an edge. We restrict our attention to monadic structures. The latter

two classes complement each other, and are essentially dual to the class MPSAT. Naturally, we do

not consider the minimization counterpart of MPSAT (where the goal is to minimize the number of

unsatisfied constraints) since it contains problems that are NP-hard for small optimum values even

for planar input-restriction [77].

Definition 33 (MPSAT) The class MPSAT consists of all NPO problems expressible as: given a

collection of FOFs over variables such that each formula Φ is a disjunction of 1

minterms, find a truth assignment of weight at most maximizing the total weight of the FOFs in

that are satisfied.


Definition 34 (TMAX) The class TMAX consists ofNPO problems expressible as: given a collec-

tion of FOFs over variables such that each formulaΦ is a disjunction of 1 negative

minterms, find a maximum weight truth assignment that satisfies each FOF in .

Definition 35 (TMIN]) The class TMIN consists of NPO problems expressible as: given a collec-

tion of FOFs over variables such that each formulaΦ is a disjunction of 1 positive

minterms, find a minimum weight truth assignment that satisfies each FOF in .

Observe that these classes are fairly strong generalizations of existing syntactic classes, in

that they allow each constraint to depend on an unbounded number of variables and also allow

each minterm to be of unbounded size. In Section 3.7, we further generalize these definitions to

problems over multi-valued domains allowing Min-Max objective functions, thereby obtaining the

class MINMAX. This class captures problems such as multiprocessor scheduling.

We now endow these definitions with a restriction on the structure of the input formulas. This

restriction is formulated in terms of incidence graphs. The classes of interest are those where the

instances are restricted to having planar incidence graphs.

Definition 36 (Incidence Graph) Let be an instance of a problem Π in MPSAT, TMAX, TMIN,

or MINMAX. Then, the incidence graph of , denoted , is defined as follows:

has a -vertex for each variable, and an -vertex for each FOF;

for each FOF Φ , and each variable in Φ, there is an edge between the -vertex for

and the -vertex for Φ.

Definition 37 (Planar Input-Restructions) Let be any syntactic optimization class. The class

PLANAR is the class restricted to instances with planar incidence graphs.

3.2.1 Main Results and Implications

Our main results are summarized in the following proposition and brief proof sketches are given in

the following sections. A similar result can be obtained for the class -OUTERPLANAR MINMAX

defined in Section 3.7; we omit the detailed exposition of this and other extensions.


Proposition 1 Each of PLANAR MPSAT, PLANAR TMAX, and PLANAR TMIN is contained in

PTAS.

Some representative applications of this proposition are listed below.

PLANAR MAX SAT belongs to PLANAR MPSAT and hence has a PTAS. This is a new result.

We believe that no PTAS was known earlier for any class of satisfiability problems involving

constraints with an unbounded number of variables. The most general previous result, due to

Hunt et al [54], was for formulas of bounded-arity drawn from a fixed family.

The knapsack problem is in PLANAR MPSAT. A variable encodes the placement of an

item in the knapsack, and a formula encodes the profit obtained by placing an item in the

knapsack. The incidence graph is a matching and hence planar. This explains the PTAS [94]

for knapsack.

The weighted MIS problem in planar graphs is in PLANAR TMAX. Vertices are encoded by

variables, and for each edge there is a formula Φ . For a planar graph ,

the resulting incidence graph is also planar. This explains the PTAS [13] for planar weighted

MIS.

The minimum weighted vertex cover (WVC) problem in planar graphs is in PLANAR TMIN.

Vertices are encoded by variables, and for each edge there is a formulaΦ .

For a planar input graphs, the incidence graph is also planar. This explains the PTAS [13] for

planar WVC.

Multiprocessor scheduling on machines, for any fixed , is in OUTERPLANAR MINMAX

with an empty set of constraints. This explains the PTAS [47] for multiprocessor scheduling.

Consider the problem of finding the maximum induced subgraph satisfying a hereditary

property Π determined by components [24]. For planar graphs, this can be placed in the class

PLANAR TMAX by an approximation-preserving reduction such as the E-reduction [67]. The

idea is to first apply planar separator theorem to delete vertices such that the largest


component size is only 1 . This explains the PTAS [24] for such induced subgraph

problems.

We can show that almost every known problem in PTAS can be placed in our framework

in a like manner. There are three notable exceptions: bin packing [96, 56], some “dense” NP-

hard problems [5], and planar-graph TSP [48]. We present one possible explanation for this

incompleteness in the first two cases.

Consider first the classical bin packing problem. Vega and Lueker [96] show that there is a

bin packing algorithm that, for any 0, achieves a packing into at most 1 OPT 1 bins

in polynomial time; in fact, Karmakar and Karp [56] give an algorithm that achieves a bound

of OPT OPT , implying a fully polynomial approximation scheme (FPTAS). While this may

appear to be a PTAS, observe that via a reduction from the partition problem [41] it is easy to show

that bin packing cannot be approximated within a ratio better than 1 5 unless P NP. The apparent

contradiction is resolved by noting the crucial role played by the additive error of 1. Their scheme is

in fact an asymptotic PTAS [83]. The asymptotic approximation ratio of an algorithm is defined

as

inf1 : for some for all

such that

where is the performance ratio for an instance [41]. The point is that while bin packing

has an asymptotic PTAS, it does not belong to PTAS and hence is ineligible for consideration

in our framework. In fact, Karmakar and Karp [56] give an algorithm that achieves a bound of

OPT OPT , implying a fully polynomial approximation scheme (FPTAS). But bin packing is

strongly NP-complete and hence cannot have an FPTAS [41]. This apparent contradiction can also

be resolved by noting that the additive error could be relatively large unless one assumes that the

optimum value is some suitably fast growing function of the input size, i.e., the Karmakar and Karp

result is actually an asymptotic FPTAS [83].

Consider now the recent results due to Arora, Karger, and Karpinski [5] where they provide a

unified scheme to obtain PTAS for dense instances of certain NP-hard optimization problems where,

informally speaking, “dense” means that the optimal solution value is within a constant factor of the


maximum possible value of a solution. For instance, they show that MAX CUT has a PTAS if the

input instance contains Ω 2 edges. Specifically, for any 0, they obtain in polynomial time a

solution of size OPT 2. Consequently, restricting to problem instances such that the optimum

value is Ω 2 , such a result translates into a PTAS scheme for MAX CUT.

In other words, we contend that these approximation schemes are possible because of the

re-interpretation of the additive error term as a relative error term when the optimum value is

appropriately constrained. These observations should however point to the delicacy of the arguments

that may be required to make further progress on the questions raised in our work.

However, there remain a couple of notable exceptions. Arora et al [5] provide a PTAS for

dense graph bisection even when the optimum value is small; clearly, this remains unexplained

by the preceding discussion. Similarly, we are also unable to capture in any obvious manner the

recent PTAS results for planar-graph TSP problem [48] and the Euclidean TSP problem [3]. It is

possible that this is only a technical difficulty, since, for instance, the PTAS result of Grigni et al

uses techniques that bear some similarity to the ones used by Baker and in our work. In any case,

our contention here is not that this framework exhaustively captures all problems in PTAS, but

merely that it provides a unified framework for explaining known membership in PTAS, and for

obtaining new results of this type. Such a unification raises the possibility that a generalization of

this framework, or some variant thereof, may yield a syntactic characterization under closure for

PTAS.

3.2.2 Possibility of Extension of Classes

In this section, we discuss the potential for extending our framework to more general classes that

retain membership in PTAS. There are two natural directions for generalizing the syntactic classes

wherein our results apply. One possibility is to relax some constraints in the logical definition of the

optimization classes. So, for instance, it may seem natural to relax the restriction of 1 minterms

in the class MPSAT by considering CNF representations of the formulas (or oracle representations,

for that matter). But this definition is immediately general enough, even with an input restriction

to trees, to capture 3-SAT as a special case, and hence would not permit a PTAS. Similarly, one

could considering relaxing the constraint enforcing strictly negative minterms in the class TMAX


and strictly positive minterms in the class TMIN. But once again, relaxing this immediately rules

out the possibility of PTAS in a strong sense since then we can encode PLANAR 3-SAT. However,

our results for MPSAT, TMAX, and TMIN generalize to non-boolean domains, i.e., they hold even

when the variables take on a polynomially-bounded number of distinct values. Furthermore, they

hold even when the notion of constraints is relaxed to functions over non-boolean domains modulo

some restrictions. Our techniques can also be used to obtain PTAS for certain classes of integer

linear programs. We omit a formal exposition of these extensions.

The other direction for extending our results is by relaxing the planarity restriction. In the final

version, we will describe an extension our results to -near-planar graphs provided that the optimum

value is proportional to the number of vertices in the incidence graph. (This is similar to the results of

Hunt et al [54] for specific MAX SNP problems which clearly satisfy the restriction on the optimum

value.) The key idea is to start with a near-planar embedding and replace each edge crossing with

a Lichtenstein gadget [77] that introduces auxiliary formulas. In other words, problems on such

graphs have an approximation-preserving reduction to planar graphs. An interesting direction for

future work is: are there more general classes of graphs to which our results can be extended? One

possibility is to consider families of graphs with excluded minors.

3.2.3 Overview of Proof Techniques

All known PTAS results for planar graph problems employ the following paradigm: find a separator

of the planar graph such that the problem is either optimally or approximately solvable on

the components resulting from the deletion of the separator; show that the separator contributes

only a small fraction of the optimal solution, obtain a solution by combining the solutions on the

components. Some earlier results [80, 24] were based on finding small separators that yielded

small components. Baker [13] observed that many NP-hard optimization problems can be solved

optimally in polynomial time on 1 -outerplanar graphs, and that separators can be found that

yield 1 -outerplanar components containing sufficient information about the optimal solution to

permit a PTAS.

Our approach extends Baker’s paradigm by computing either an exact solution or a PTAS

for generic syntactic problems on outerplanar graphs, and then extending this to planar graphs.


While Baker’s paradigm employs dynamic programming on 1 -outerplanar explicitly based on

a planar embedding, our approach, on the other hand, is to use a more structured and uniform

representation of outerplanar graphs in terms of a tree-decomposition [21]. In our abstraction,

dynamic programming combined with rounding techniques can be used to solve optimally or

approximately the proposed syntactic problems when the incidence graphs are restricted to being

1 -outerplanar. The syntactic structure of the problem is then used to extend the solution to the

original planar graph by either using a second-level of dynamic programming, or a pigeon-hole

argument similar to Baker’s.

3.3 Outerplanar Graphs and Tree Decompositions

Our constructions of PTAS for the planar versions of these complexity classes depends on the

strengthening of planar separator theorem by Bodlaender [21] for 1 -outerplanar graphs.

Definition 38 ( -Outerplanar Graph) A 1-outerplanar graph, or simply an outerplanar graph, is a

planar graph that has an embedding on the plane with all vertices appearing on the outer face; a

-outerplanar graph is a planar graph that has an embedding on the plane in which deleting all

vertices on the outer face yields a 1 -outerplanar graph.

Observe that the outer face of a 1-outerplanar graph need not be a simple cycle; for example,

a tree is outerplanar and two cycles meeting at a single vertex is also outerplanar. We describe a

useful structural property of outerplanar graphs based on the notion of treewidth due to Robertson

and Seymour [93].

Definition 39 (Tree Decomposition) A tree decomposition of a graph is a 2-tuple

such that:

for each , there exists an with ;

for all , there exists an with ;

and, for all , the set induces a subtree.


Definition 40 (Width and Treewidth) The width of a tree-decomposition

of a graph is max 1 The treewidth of a graph is the minimum width

over all tree-decompositions.

For any node in the decomposition tree, define the graph as the subgraph of the graph

induced by the set of vertices consisting of the vertices stored at all the nodes in the subtree

rooted at . Thus, for the root , we have and .

Proposition 2 A tree decomposition of a graph may be viewed as a separator tree. In particular,

at the root with children and ,

1. , and

2. there are no edges from to in .

Further, the same property holds recursively at each node in the tree.

Lemma 6 ([21]) A -outerplanar graph has treewidth at most 3 1.

Definition 41 (Nice Tree Decomposition) A tree decomposition is

nice if one can choose a root in the tree such that:

is a binary tree,

if is a leaf, then 1 (START node);

if has two children and , then (JOIN node);

and, if has one child , then either

– there exists a vertex with (FORGET node), or

– there exists a vertex with (INTRODUCE node).

We use the following known results: there is always a nice tree decomposition of optimal

width [21]; and, there is a linear time algorithm to a construct tree-decomposition of graphs with

1 treewidth [22].


3.4 Planar MPSAT

Let denote the weight associated with the variable and denote the profit associated with

the formula ; furthermore, let denote the weight threshold and let denote the sum of all the

profits.

We begin by sketching a pseudo-polynomial time algorithm for the case of 1 -outerplanar

graphs. The first step is to compute a nice tree decomposition for a given 1 -outerplanar graph

; recall that it has treewidth 1 . We need some definitions.

Definition 42 (Tuple) A tuple over a set of variables 1 and a set of formulas 1 is

a vector of the form

1 1

where each 0 1 and each is either a minterm that occurs in formula or is NULL.

Definition 43 (Consistent Tuple) A tuple

1 1

over a set of variables 1 and formulas 1 is called a consistent tuple if whenever a

variable occurs in a minterm in , then 1 if occurs as a positive literal in , and

0 otherwise.

In our algorithms, we maintain the following 2-dimensional structure at each node in the tree

.

Definition 44 ( VLIST) For a node in the decomposition tree, VLIST denotes the minimum

weight needed to achieve a profit of in the MPSAT subproblem represented by the subgraph

when the variables and formulas at node are set in accordance with the tuple .

Definition 45 ( EXT) Given a tuple and a variable or formula , the extension set EXT

consists of all tuples that are extensions of with each possible value that can be assigned to .


Definition 46 ( PROJ) Given a tuple and a variable or formula , the projection PROJ is the

tuple obtained from by dropping the component corresponding to .

Definition 47 ( WT and PT ) Given a tuple , WT denotes the weight of variables set to true in ,

and PT denotes the total profit associated with formulas appearing in with non-NULL minterms.

We now show that the VLIST array can be efficiently computed at the root of a nice treewidth

decomposition tree using a bottom-up approach. The entries in the VLIST array at each node are

initialized to . At the leaves, the array can be trivially computed in polynomial time. To

compute the array at an intermediate node such that it has already been computed for its children,

we analyze three cases:

at a JOIN node with children and , we compute

VLIST min0

VLIST PT VLIST WT

at a FORGET node with a child such that , we compute

VLIST minEXT

VLIST

finally, at an INTRODUCE node with a child such that , we consider two

subcases.

If corresponds to a variable then

VLIST

VLIST PROJ if is consistent and is false in

VLIST PROJ if is consistent and is true in

otherwise

On the other hand, if corresponds to a formula then


VLIST

VLIST PROJ

if is consistent and contains a non-NULL minterm for

VLIST PROJ

if is consistent and contains a NULL minterm for

otherwise

In all cases, the computation of VLIST takes polynomial time, and we perform this computation

exactly once at each node. Hence, the VLIST computation at the root of takes only polynomial

time. Finally, observe that by definition of VLIST, the largest value occurring in VLIST is

the desired optimal solution. The proof of correctness of this computation, as well as a bound of1 2 on the running time, is omitted. By using a standard technique of rounding the lower

order bits, we can obtain a PTAS from this pseudo-polynomial time algorithm. The details are

omitted; see for example [86]. We obtain the following results.

Lemma 7 The class 1 -OUTERPLANAR MPSAT is contained in PTAS.

Corollary 1 1 -OUTERPLANAR MPSAT is in P when the formulas have uniform profits.

Thus satisfiability problems such as MAX 3-SAT become polynomial-time solvable on this

subclass of planar graphs. Consider now the following theorem.

Theorem 20 The class PLANAR MPSAT is in PTAS.

Proof: This is established by combining Baker-style arguments with a dynamic programming

scheme. Consider the input graph and assume that it is -outerplanar for , where is

the number of variables and is the number of formulas in the collection . Divide the vertices

into levels, called 1 , such that corresponds to the outer face and each level is the

outer face obtained by removing the levels 1.

Fix any optimal truth assignment, and let denote the weight of the clauses satisfied in level

by this assignment. Partition the levels 1 into 1 groups, 0 , where group


is the union of the levels whose index is congruent to 3 , 3 1, or 3 2 modulo , where

3 1 . By pigeon-holing, there exists a group such that

OPT1

This special group may be identified by trying all possible choices of , and picking the one

which yields the best solution. Having fixed the choice of , delete all vertices in the levels whose

index is congruent to 3 1 modulo , thereby separating the graph into a collection of disjoint

1 -outerplanar graphs, say 1 2 , such that the total optimal value on this collection

is at least 1 3 OPT.

For a given weight threshold , we know how to obtain a pseudo-polynomial time algorithm

for any . But we need to find an “optimal partitioning” of the total weight into sub-thresholds

1 2 , where is the weight threshold allocated to . To this end, we begin by

performing the VLIST computation for each . Once this two-dimensional array is computed, we

can extract from it a one-dimensional profit array 0 for each , such that denotes the

minimum weight needed to achieve profit in the instance encoded by .

Next, we create a complete binary tree whose leaves correspond to 1 2 and an

intermediate node corresponds to the union of the graphs represented by its children. Thus, the root

corresponds to 1 . This can now be used to iteratively compute the profit arrays associated

with any non-leaf vertex. Consider a non-leaf vertex such that we already have the profit arrays

for its two children and . The array at can be computed as follows:

0 0 min

We omit the formal proof of correctness of this computation and of the total running time bound

as a polynomial in and . Thus, we have a pseudo-polynomial time approximation scheme for

the constrained optimization problem. The final step is to use the standard lower-order bit rounding

technique to obtain a PTAS. We omit the details (refer to Papadimitriou [86]).


3.5 Planar TMAX

We sketch the PTAS for PLANAR TMAX. As before, is the weight of variable ; since all

constraints must be satisfied, we do not associate any profits with the formulas. We begin with the

following lemma.

Lemma 8 The class 1 -OUTERPLANAR TMAX is contained in P.

In the proof sketch for this lemma, we will define VLIST as a one-dimensional array.

Definition 48 ( VLIST) For a node in the decomposition tree, VLIST denotes the maximum

weight achievable in the TMAX subproblem represented by the subgraph when the variables

and formulas at node are set in accordance with the tuple .

The entries in the VLIST array at each node are initialized to . In the rest of this section, we

will assume that a tuple does not contain a NULL term for any formula, i.e., we consider only those

tuples that satisfy all their formulas.

We briefly sketch the computation of the VLIST array at the root of a nice treewidth decomposition

tree using a bottom-up approach. At the leaves, the array can be trivially computed in polynomial

time. To compute the array at an intermediate node where it has already been computed for its

children, we consider three cases:

at a JOIN node with children and , we compute

VLIST VLIST WT

at a FORGET node with a child such that , we compute

VLIST maxEXT

VLIST

finally, at an INTRODUCE node with a child such that , we consider two

subcases.


If corresponds to a variable then

VLIST

VLIST PROJ if is consistent and is false in

VLIST PROJ if is consistent and is true in

otherwise

On the other hand, if corresponds to a formula then

VLISTVLIST PROJ if is consistent

otherwise

We omit the proof of correctness of this computation and the running time bound of 1 .

A careful look at this computation reveals that we did not really use the negativity constraint on the

minterms and thus we can conclude a stronger result.

Corollary 2 The class 1 OUTERPLANAR TMAX without negativity constraints on the minterms

is in P.

We now sketch the extension to planar graphs leading to the following theorem.

Theorem 21 The class PLANAR TMAX is contained in PTAS.

Proof Sketch : The idea is to remove a collection of layers whose deletion leaves an instance with

optimal value at least 1 1 OPT and which is a collection of -outerplanar graphs, say

1 . Fix any optimal assignment and let denote the weight of the variables set to true in

level . Partition the levels 1 into 1 groups, 0 , where group is the union

of the levels whose index is congruent to 2 or 2 1 modulo , where 2 1 . By

pigeon-holing, there exists a group such that

OPT1

This special group may be identified by trying all possible choices of , and picking the one that

yields the best solution. Having fixed the choice of , delete all variable vertices in the levels


represented in . Compute the optimal solutions in each of the resulting 2 1 -outerplanar

graphs (recall that the formula vertices do not have any edges between them). Since these graphs

do not share any vertices, we can compute their optimal solutions independently and combine these

solutions without any conflict. The combined solution is finally extended by setting all the variables

in the special group to FALSE and the remaining variables according to the value assigned by

the above solutions. It is easy to see that due to the negative minterm property, this scheme of

composing the optimal solution results in a solution satisfying all the constraints. And finally, the

weight of the solution is at least 1 1 1 OPT.

3.6 Planar TMIN

An algorithm similar to the one in Section 3.5 can be used to establish that 1 -OUTERPLANAR

TMIN is in P and also that 1 -OUTERPLANAR TMIN without positivity constraints on the minterms

lies in P. However, the scheme needed to create a PTAS for planar graph is slightly different; it

is a variant of the scheme used by Baker [13] for the minimum vertex cover problem. The idea is

to consider groups of consecutive layers such that the last two layers of any group are identical

to the first two layers of the next group. If we compute solutions for the -outerplanar graphs

defined by these layers and combine them by setting a variable to TRUE whenever it is set to TRUE

in any solution and FALSE otherwise, it is easy to see that such a solution must indeed satisfy all

the constraints because each constraint vertex along with its constituent variable vertices must be

completely contained in one of the groups. Now, if we fix any optimal assignment, there exists a

way of defining a shift of these layers such that the total weight of OPT in the overlapping layers is

only OPT . Hence, we have a PTAS.

3.7 1 -Outerplanar MINMAX

In this section we generalize our syntactic classes to capture problems over multi-valued domains

where the objective is to have a “balanced solution” satisfying a given set of constraints. However,

our results hold only for 1 -outerplanar graphs – a class general enough to capture problems such


as multiprocessor scheduling as a special case.

Definition 49 ( -ary term)

A -ary term over a set of variables is a constraint of the form 1 1 where

0 and is a constant. An assignment satisfies the term if and only if its projection on the

variables in the term coincides with the specified values.

Definition 50 ( -ary formula)

A -ary formula over a set of variables is a set of -ary terms such that the formula is satisfied by

an assignment if and only if it satisfies at least one term in the formula.

Definition 51 (Variable Class)

Given an assignment , a variable class denotes the set of variables with value in assignment .

We define a new class as follows.

Definition 52 (MINMAX)

The class MINMAX consists of NPO problems expressible as: given a collection of FOFs over

variables such that each formula Φ has 1 -ary terms, find an assignment which

minimizes the maximum-weight variable class.

We briefly sketch the idea used to obtain a PTAS for 1 -OOUTERPLANAR MINMAX. As we

did before, for 1 -outerplanar graphs we perform a bottom-up VLIST computation on a nice tree-

decomposition. The VLIST array is a 2-dimensional array – the array entry VLIST 0 1

stores a YES/NO answer to the question of whether there is a feasible solution which respects the

constraint set and the variable assignment specified by the tuple, and partitions the variable weights

into 0 1 ; where 0 and 1 WT . This computation can be

performed in time 1 , since is a constant. At the root, we simply choose the optimal

feasible solution over all tuples. Hence, we have a pseudo-polynomial time algorithm. But observe

that OPT 1 and therefore, once again using the lower order bit rounding techniques,

we can convert this into a PTAS result.

Chapter 4

The Approximability of Constraint

Maximization Problems

80

CHAPTER 4. THE APPROXIMABILITY OF CONSTRAINT MAXIMIZATION PROBLEMS81

4.1 Introduction

Thus far, our focus has been to identify the structure that underlies all problems in a given approx-

imation class. But if we consider an approximation class such as APX, we observe that not all

problems in this class are equally hard. In particular, APX contains problems in P which we can

solve optimally and problems such as MAX 3-SAT which can’t be approximated beyond a certain

constant factor. At this juncture, we would like to ask if we can characterize which optimization

problems are easy and which ones hard; and to ask if there are any characteristic features, specifc

only to hard problems, that can be isolated. Of course, no complete characterization is possible:

Rice’s theorem allows one to disguise an optimization problem so cleverly that it would be undecid-

able to determine if a given problem is NP-hard or polynomial time solvable. Even if the problem is

presented in its simplest form — it may be the case that the answer need not be "easy" or "NP-hard"

— this is established by a theorem of Ladner [75].

In the presence of such barriers one is forced to weaken one’s goals and focus one’s attention onto

a restricted subclass of NP Optimization problems in the hope that some features of hard problems

can be isolated from this subclass. Our choice of the appropriate restriction comes from the work of

Schaefer [95] who carried out an analogous investigation in the case of decision problems. Schaefer

considered a restriction of NP that he called “satisfiability problems” and successfully characterized

every problem in this (infinite) class as being easy (polynomial time decidable) or hard (NP-hard).

He refers to this as a “dichotomy theorem”, since it partitions the class of problems studied into

two polynomial time equivalent classes. A further study along these lines — looking for other

dichotomic classes within NP — was carried out more recently by Feder and Vardi [35].

Schaefer’s work is central to our technical results. We sketch his result in some detail here; a

formal presentation may be found in Section 4.2. Schaefer considers what he calls “generalized

satisfiability problems”. A typical instance to a typical problem consists of boolean variables

with constraints on them. The constraints are of finite arity and come from a finite collection of

templates, , which define the satisfaction problem, which we will call SAT( ). For every such

collection he investigates the complexity of determining if the given instance is satisfiable, i.e.,

This chapter is based on joint work with Madhu Sudan [70].


is there an assignment to the variables satisfying all constraints. He enumerates six classes of

constraint sets for which the satisfiability problem SAT( ) is decidable in polynomial time. These

classes include 2CNF formulae, satisfiability of linear equations modulo two and satisfiability of

Horn clauses (and some other simple classes). He shows that whenever a family contains members

not strictly in the six cases he considers, then the satisfibility problem is NP hard!

We generalize Schaefer’s characterization to optimization problems. We define a class of

optimization problems called MAXCSP which is a subclass of MAX SNP and forms what we feel

is a combinatorial core of MAX SNP. As in Schaefer’s case, a problem of this class is described by

a constraint set . An instance of this problem is given by boolean variables and constraints,

and the objective is to find an assignment satisfying the maximum number of constraints. It turns

out that MAXCSP contains most of the well-known MAX SNP problems, such as MAX CUT,

MAX 2SAT and MAX 3SAT. We show that every problem in MAXCSP is either solvable exactly

in polynomial time or MAX SNP-hard. We are able to characterize exactly the structure of the easy

problems. Our characterization of easy problems in this class shows that - MIN CUT problem is

essentially the only easy problem in this class. Barring this problem, every other problem that we

study in this class, is complete for the class (and hence for MAX SNP). In particular, there are no

problems which possess a PTAS (without possessing an exact optimization algorithm). Our result

provides a new insight into questions such as : Why do all the known MAX SNP problems that are

NP-hard, also turn out to be MAX SNP-hard? What is the unifying element among these problems?

The remainder of this chapter is organized as follows. Section 4.2 introduces the definitions

and notation needed to describe our work. A formal statement and an overview of our main result

is given in Section 4.3. Sections 4.4 and 4.5 are devoted to proving the main result, that is, the

classification theorem for MAXCSP. In Section 4.6, we show that the MAXCSP version of the

problems characterized as hard by Schaefer, are MAX SNP-hard at the gap location 1. Finally, in

Section 4.7, we show how to eliminate the replication assumption from Schaefer’s proof.

4.2 Preliminaries

We start with the definition of a constraint.


Definition 53 [Constraint] A constraint is a function : 0 1 0 1 . We say is satisfied by

an assignment 0 1 if 1. We refer to as the arity of the constraint . A constraint

with no satisfying assignments is called unsatisfiable.

Often we apply a constraint of arity to a subset of variables from a larger set. In such

cases we think of as a constraint on the larger set of variables.

Definition 54 [Constraint Application] Given boolean variables 1 and a constraint

of arity , and indices 1 1 , the pair 1 is referred to as an

application of the constraint to 1 . An assignment for 1 and

0 1 satisfies the application if 1 1.

Definition 55 [Constraint Set] A constraint set 1 is a finite collection of constraints.

Definition 56 [Constraint Satisfaction Problem (MAXCSP )] Given a constraint set , the

constraint satisfaction problem MAXCSP is defined as follows :

INPUT : A collection of constraint applications of the form 1 1, on

boolean variables 1 2 where and is the arity of .

OBJECTIVE : Find a boolean assignment to ’s so as to maximize the number of applications

of the constraints 1 that are satisfied by the assignment.

Notice that the above definition gives a new optimization problem for every family . Schaefer’s

class of decision problems SAT( ) can be described in terms of the above as: The members of

SAT( ) are all the instances of MAXCSP whose optimum equals the number of applied

constraints (i.e., all constraints are satisfiable). Schaefer’s dichotomy theorem essentially shows

that the only families for which SAT( ) is in P, is if all constraints in are either satisfied

by the all zeroes assignment, the all ones assignment or all constraints are linear constraints over

GF(2) or all constraints are Horn clauses. A formal statement is as below; we need some additional

definitions.

Definition 57 [Weakly Positive and Weakly Negative Functions] A function is called weakly posi-

tive (weakly negative) if it may be expressed as a CNF formula such that each clause has at most

one negated (unnegated) variable.


Definition 58 [Affine] A function is said to be affine if it may be expressed as a system of linear

equations of the form 1 0 and 1 1 (i.e. it evaluates to one iff the input variables

satisfy the given equation system); the addition operation being modulo 2.

Theorem 22 ([95]) Let be a finite set of boolean functions. Then SAT is always either in P

or NP-hard. Furthermore, it is in P if and only if one of the following conditions is true:

1. Every is 0-valid.


3. Every is weakly positive.

4. Every is weakly negative.

5. Every is affine.

6. Every is bijunctive (i.e. expressible as a CNF formula with at most 2 literals per

clause).

Our main result asserts that a dichotomy holds for MAXCSP as well. However the char-

acterization of the easy functions and the hard ones is quite different. (In particular, many more

constraint sets are hard now.) In order to describe our result fully we need some more definitions.

Definition 59 Given a constraint set , the constraint set is the set of constraints in which

are satisfiable.

It is easy to see that for any constraint set , an instance of MAXCSP can be mapped to

an instance of MAXCSP , such that the objective function value is preserved on each input

assignment. Hence our characterizations will essentially be characterizations of MAXCSP .

Definition 60 [ -valid function] For 0 1 , a function of arity is called -valid if it is

satisfied by the assignment .


Definition 61 [Minterm] Given a function on variables 1 , a conjunction of literals

drawn from these variables, say 1 1 , is called a minterm of if it satisfies the

following properties:

1. Any assignment 1 , which satisfies 1 1 and 1 0

satisfies .

2. The collection is minimal with respect to property (1).

Definition 62 [Positive and Negative Minterms] A minterm of which consists only of unnegated

variables is called a positive minterm. A minterm which consists only of negated variables is called

a negative minterm.

Definition 63 [2-Monotone Function] A function is called 2-monotone if it has at most two

minterms such that at most one of them is positive and at most one is negative.

4.3 Overview of the Main Result

Our main result is as stated below.

Theorem 23 The problemMAXCSP is always either in P or isMAX SNP-hard. Furthermore,

it is in P if and only if one of the following conditions is true:



3. Every is 2-monotone.

This theorem follows from Lemmas 9,10,11 and 23. A second result that follows easily as

a consequence of Schaefer’s result and our notion of approximation preserving reductions is the

following:

Theorem 24 For every constraint set either SAT( ) is easy to decide, or there exists 0

such that it is NP-hard to distinguish satisfiable instances of SAT , from instances where 1

fraction of the constraints are not satisfiable.


Schaefer’s result characterizes which function families are easy to decide and which ones are

hard.

4.3.1 Discussion

The main feature of Theorem 25 is that the family of constraint sets which lead to hard problems is

significantly larger than in Schaefer’s case. This is not surprising given that problems such as 2SAT

and Linear systems over GF(2) are easy problems for the decision version and the maximization

versions are known to be hard, even to approximate [87, 4]. Nevertheless, the set of problems

that are shown to be easy is extremely small. MAXCSP for which is 0-valid or 1-valid is

really a trivial problem; leaving only the class of 2-monotone functions as somewhat interesting.

But the class of functions with such properties seems to be really small and maximum flow or -

min cut appear to be the only natural optimization problems with this property. Thus, we feel, that

the correct way to interpret Theorem 25 is to think of it as saying that every constraint satisfaction

problem is either solvable by a maximum flow computation or it is MAX SNP-hard.

Another interesting feature of the above result is that the dichotomy holds for two different

properties — the complexity and the approximability — simultaneously. Easy problems are easy to

compute exactly and the hard ones are hard to even approximate. The middle regime — problems

that are easy to approximate but hard to compute exactly — are ruled out. This may be somewhat

surprising initially, but becomes inevitable once the form of approximation preserving reductions

we use here becomes clear. Essentially all reductions we use are exactly those that might be used

for exact optimization problems. In fact the ease with which these reductions apply is the reason

why Theorem 24 falls out easily from this work.

The technical aspects of the proof of the dichotomy theorem may be of some independent

interest. In order to prove such a theorem, one needs to find succinct characterizations of what

makes a function, say 2-monotone, as well as a succinct proof when a function is not 2-monotone. We

find such a characterization, in Lemma 17, which turns out to be useful in establishing Theorem 25.

One technical nit-picky point that we face in this study is the role of constants and repe-

titions in MAXCSP. In particular, if , should 1 0 given by 1 0 2

0 2 also be considered a member of the constraint set? Similarly should 1 2


1 1 2 2 be considered a member of the constraint set. Allowing for these repeti-

tions and constants makes the analysis much easier; however they may change the complexion of

the problems significantly. For instance given a set of linear equalities of the form 0, it

is trivial to find an assignment which satisfies all the equations — namely the all 0’s assignment.

However once one is allowed to fix some variables to the constant 1, the problem no longer remains

easy. In our presentation, we initially assume we can use constants and repetitions to make our

analysis simpler. Later we remove the assumptions — and Theorem 25 and Theorem 24 is shown

without the use of any constants or repetition. In fact, in the process we remove a minor irritant

from Schaefer’s proof which actually needed to use repetitions.

4.3.2 Related Work

Theorem 25 was independently discovered by Creignou [26]. Here we clarify the main points of

difference between this work and that of [26]. In [26], the “easy" problems are characterized by

a graph-theoretic representation (this is possible since the functions involved in the easy side can

be expressed as CNF formulas such that each clause is implicative, that is of the form

) and the proof uses graph-theoretic ideas. Our result are stated and established

via techniques in the more general context of constraint satisfaction. The proof might be adapted

to problems over other domains (larger than the boolean one), whereas it seems unlikely that one

could extend the proof of [26] in such a way.

Additionally, technical aspects of the proof given here may be of some independent interest.

Particularly, the notion of -implementation defined here gives a clear way to translate all the

hardness results shown here into hardness of approximation results. Also, the fact that we do

not need to use repetition of variables in functions, to obtain hardness results is another technical

improvement on previous results, including that of [95].

4.4 Polynomial Time Solvability

From this section onwards we omit the notation and assume we have a constraint such that

.


Lemma 9 The problem MAXCSP is in P if each is 0-valid.

Proof: Set each variable to zero; this satisfies all the constraints.

Lemma 10 The problem MAXCSP is in P if each is 1-valid.

Proof: Set each variable to one; this satisfies all the constraints.

Lemma 11 The problem MAXCSP is in P if each is a 2-monotone function.

Proof: We reduce the problem of finding the maximum number of satisfiable constraints to the

problem of finding the minimum number of unsatisfied constraints. This problem, in turn, reduces

to the problem of finding - min-cut in directed graphs. 2-monotone constraints have the following

possible forms : (a) 1 2 , (b) 1 2 , and (c) 1 2 1 2 where 1.

Construct a directed graph with two special nodes and and a vertex corresponding

to each variable in the input instance. Let denote an integer larger than the total number of

constraints. Now we proceed as follows for each of the above classes of constraints :

For a constraint of the form (a), create a new node and add an edge from each to

of cost and a unit cost edge from to .

For a constraint of the form (b), create a new node and add an edge of cost from

to each and an edge from to of unit cost.

Finally, for a constraint of the form (c), we create two nodes and and connect to

1 2 and connect to 1 2 as described above and replace the unit cost edges

from and to by a unit cost edge from to .

Using the correspondence between cuts and assignments which places vertices corresponding

to true variables on the side of the cut, we find that the cost of a minimum cut separating from

, equals the minimum number of constraints that can be left unsatisfied.

The next lemma shows that - min-cut problem with polynomially bounded integral weights

is in MAXCSP for some 2-monotone constraint set . Since the previous lemma shows how


to solve MAXCSP for any 2-monotone constraint set by reduction to - min-cut problem,

it seems that - min-cut problem is the hardest (and perhaps the only) interesting problem in

MAXCSP .

Lemma 12 The - min-cut problemwith polynomially bounded integralweights is inMAXCSP

for some 2-monotone constraint set .

Proof: Let be a family of three functions 1 2 3 such that 1 ¯ 2

¯ and 3 .

Now given an instance to - min-cut problem, we construct an instance of

MAXCSP on variables 1 2 where corresponds to the vertex :

For each edge with weight , we have copies of the constraint 1 .

For each edge with weight , we have copies of the constraint 3 .

For each edge with weight and such that , we have copies of

the constraint 2 .

Given a solution to this instance of MAXCSP , we construct an - cut by placing the vertices

corresponding to the false variables on the -side of the cut and the remaining on the -side of the cut.

It is easy to verify that an edge contributes to the cut iff its corresponding constraint is unsatisfied.

Hence the optimal MAXCSP solution and the optimal - min-cut solution coincides.

4.5 Proof ofMAX SNP-Hardness

In this section we prove that a constraint set which is not entirely 0-valid or entirely 1-valid or

entirely 2-monotone gives a MAX SNP-hard problem. The main MAX SNP-hard problem which

we reduce to any of these new ones is the MAX CUT problem shown to be MAX SNP hard by

Papadimitriou and Yannakakis [87]. Initially we consider the case where we are essentially allowed

to repeat variables and set some variables to true or false. This provides a relatively painless proof

that if a function is not 2-monotone, then it provides a MAX SNP hard problem. We then use


the availability of functions that are not 0-valid or 1-valid to implement constraints which force

variables to be 1 and 0 respectively, as well as to force variables to be equal. This eventually allows

us to use the hardness lemma. We first start with some notation.

4.5.1 Notation

Given an assignment to an underlying set of variables, denotes the set of positions corre-

sponding to variables set to zero and denotes the set of positions corresponding to variables

set to one. More formally, given an assignment 1 2 to 1 2 , where 0 1 ,

we have 0 and 1 . The notation 0 denotes the set of

all assignments such that and similarly, 1 denotes the set of all assignments

such that .

Definition 64 [Unary Functions] The functions and ¯ are called unary

functions.

Definition 65 [XOR and REP Functions] The function is called the XOR

function and its complement function, namely , is called the REP function.

Definition 66 [ -closed Function] A function is called -closed (or complementation-closed) if

for all assignments , ¯ .

Definition 67 [ -Consistent Set] A set of positions is -consistent for a constraint iff every

assignment with all variables occupying the positions in set to value is a satisfying assignment

for .

4.5.2 -Implementations andMAX SNP-Hard Functions

We next describe the primary form of a reduction which we use to give the hardness results. As

pointed out by Papadimitriou and Yannakakis, in order to get hardness of approximation results, the

reductions used need to satisfy certain approximation preserving features. Here we show how to

implement a given function using a family of other functions , so as to be useful in approximation

preserving reductions.


Definition 68 [ -Implementation] An instance of MAXCSP over a set of variables

1 2 and 1 2 is called an -implementation of a boolean function

, where is a positive integer, iff the following conditions are satisfied:

(a) no assignment of values to and can satisfy more than constraints,

(b) for any assignment of values to such that is true, there exists an assignment of values

to such that precisely constraints are satisfied,

(c) for any assignment of values to such that is false, no assignment of values to can

satisfy more than 1 constraints, and finally

(d) for any assignment to which does not satisfy , there always exists an assignment to such

that precisely 1 constraints are satisfied.

We refer to the set as the function variables and the set as the auxiliary variables.

Thus a function 1-implements itself. We will say that MAXCSP implements function if

it -implements for some constant . The following lemma shows that the -implementations

of functions compose together. The criteria for “implementation” given above are somewhat more

stringent than used normally. While properties (1)-(3) are perhaps seen elsewhere, property (4) is

somewhat more strict, but turns out to be critical in composing implementations together.

Lemma 13 [Composition Lemma] Given two constraint sets and such that MAXCSP

can -implement a function , and MAXCSP can -implement a function , then

MAXCSP can -implement the function for some constant .

Proof: Let be the number of occurrences of a constraint involving the function in the

MAXCSP instance 1-implementing , then clearly, by replacing each occurrence of by its

2-implementation, we obtain a MAXCSP instance which -implements for

1 2 1 .

The MAX SNP-hardness of MAX CUT implies that MAXCSP XOR is MAX SNP-hard

and hence the below :


Lemma 14 If MAXCSP can implement the XOR function, then MAXCSP is MAX SNP-

hard.

Lemma 15 MAXCSP can implement theXOR function if is either the function

or ¯ or ¯ ¯ .

Proof: If , then the instance is a 3-implementation

of ; if ¯ , then the instance is a 1-implementation of ; and

finally, if ¯ ¯ , then is a 3-implementation of .

Lemma 16 MAXCSP can implement the REP function if is the function ¯ .

Proof: The instance is a 3-implementation of the function

REP.

4.5.3 Characterizing 2-Monotone Functions

In order to prove the hardness of a constraint which is not 2-monotone, we require to identify some

characteristics of such constraints. The following gives a characterization, which turns out to be

useful.

Lemma 17 [Characterization Lemma] A function is a 2-monotone function if and only if all the

following conditions are satisfied:

(a) for every satisfying assignment of , either 1 or 0 is a set of satisfying

assignments,

(b) if 1 is 1-consistent and 2 is 1-consistent for , then 1 2 is 1-consistent, and

(c) if 1 is 0-consistent and 2 is 0-consistent for , then 1 2 is 0-consistent.

Proof: We use the fact that a function can be expressed in DNF form as the sum of its minterms.

For a 2-monotone function this implies that we can express it as a sum of two terms. Every satisfying

assignment must satisfy one of the two terms and this gives Property (a). Properties (b) and (c) are

obtained from the fact that the function has at most one positive and one negative minterm.


Conversely, if a function is not 2-monotone, then it either has a minterm which is not monotone

positive or negative or it has more than one positive (or negative) minterm. In the former case, the

function will violate Property (a), and in the latter one of Properties (b) or (c).

Observe that a 2-monotone function is always either 0-valid or 1-valid or both.

4.5.4 MAX SNP-hardness of Non 2-Monotone Functions

We now use the characterization from the previous subsection to show that if one is allowed to

“force” constants or “repetition” of variables, then the presence of non-2-monotone constraint

gives hard problems. Rather than using the ability to force constants and repetitions as a binding

requirement, we use them as additional constraints to be counted as part of the objective function.

This is helpful later, when we try to remove the use of these constraints.

Lemma 18 [Hardness Lemma] If is not 2-monotone, MAXCSP REP can implement

the function XOR.

Proof: We prove this by using the Characterization Lemma for 2-monotone functions. Let

denote the arity of . If is not 2-monotone, it must violate one of the three conditions

and stated in the Characterization Lemma.

Suppose violates the property (a) above. Then for some satisfying assignment , there exist

two assignments 0 and 1 such that 0 and 1 , but 0 1 0.

Without loss of generality, we assume that 0 1 , 0 0 1 and 1 0 1 . Thus we

have the following situation :

00 0 00 0 11 1 11 1 1

0 00 0 00 0 00 0 11 1 0

1 00 0 11 1 11 1 11 1 0

2 00 0 11 1 00 0 11 1

Observe that both and are non-zero. We consider the MAXCSP REP instance

with the following set of constraints on variables 1 2 :


constraints for 1 ,

constraints REP 1 for 2 ,


constraints for 1 , and

the constraint 1 2 .

It is now easy to verify that for 1 , this instance -implements the function 1

1 if 2 1 and 1 1, otherwise. The claim now follows immediately from

Lemma 15.

Next suppose violates the property (b) above. Then there exists an unsatisfying assignment

such that sets all variables in 1 2 to 1, and at least one variable in each of 1 1 2 and

2 1 2 to be false. Consider one such unsatisfying assignment . Without loss of generality,

we have the following situation :

1

1

00 0 11 1

2

1 2

11 1 11 12

00 0 00 0 11 1

We consider the MAXCSP REP instance with the following set of constraints on

variables 1 2 :


constraints for 1 ,


constraints for 1 ,

constraints for , and finally


the constraint 1 2 where .

It is now easy to verify that for 1 , this instance -implements the function 1

1. Again, the claim now follows immediately from Lemma 15.

Finally, the case in which violates the property above, can be handled in an analogous

manner.

4.5.5 Implementing the REP Function

We now start on the goal of removing the use of the unary and replication constraints above. In

order to do so we use the fact that we have available to us functions which are not 0-valid and not

1-valid. It turns out that the case in which the same function is not 0-valid and not 1-valid and

further has the property that its behavior is closed under complementation (i.e., ¯ ) is

somewhat special. We start by analyzing this case first.

Lemma 19 [Replication Lemma] Let be a non-trivial function which is -closed and is neither

0-valid nor 1-valid. Then an instance of MAXCSP can implement the REP function.

Proof: Let denote the arity of and let 0 and 1 respectively denote the maximum number

of 0’s and 1’s in any satisfying assignment for ; clearly 0 1. Now let 1 2 2

and 1 2 2 be two disjoint sets of 2 variables each. We begin by creating an

instance of MAXCSP as follows. For each satisfying assignment , there are 2 2

constraints in such that every -variable subset of appears in place of 0’s in and every

variable subset of appears in place of 1’s in the assignment , where denotes the number

of 0’s in .

Clearly, any solution which assigns identical values to all variables in and the complementary

value to all variables in , satisfies all the constraints in . Let and respectively denote the

set of variables set to zero and one respectively. We claim that any solution which satisfies all the

constraints must satisfy either or .

To see this, assume without loss of generality that . This implies that

or else there exists a constraint in with all its input variables set to zero and is hence unsatisfied.


This in turn implies that no variable in can take value one; otherwise, there exists a constraint

with 1 1 of its inputs set to one, and is unsatisfied therefore. Finally, we can now conclude that

no variable in takes value zero; otherwise, there exists a constraint with 0 1 of its inputs

set to zero and is unsatisfied therefore. Thus, . Analogously, we could have started with

the assumption that and established . Hence an assignment satisfies all the

constraints in iff it satisfies either the condition or the condition .

We now augment the instance of MAXCSP as follows. Consider a least hamming weight

satisfying assignment for . Without loss of generality, we assume that 10 1 . Clearly then,

0 11 is not a satisfying assignment. Since is -closed, we have the following situation :

0 00 0 11 1 0

1 00 0 11 1 1

¯ 0 11 1 00 0 1¯ 1 11 1 00 0 0

Consider the constraints 1 2 1 2 and 1 2 1 2 .

If 1, then to satisfy the constraint 1 2 1 2 , we must have .

Otherwise, we have 0 and then to satisfy the constraint 1 2 1 2

we must have . In either case, the only way we can also satisfy the constraint

1 2 1 2

is by assigning an identical value. Thus these set of constraints -implements the function

where is simply the total number of constraints; all constraints can be satisfied iff

and otherwise, there exists an assignment to variables in and such that precisely 1

constraints are satisfied.


4.5.6 Implementing the Unary Functions

If the function(s) which is (are) not 0-valid and 1-valid is (are) not closed under complementation,

then they can be used to get rid of the unary constraints. This is shown in the next lemma.

Lemma 20 [Unary Lemma] Let 0 and 1 be two non-trivial functions, possibly identical, which

are not 0-valid and 1-valid respectively. Then if neither 0 nor 1 is -closed, an instance of

MAXCSP 0 1 can implement both the unary functions and .

Proof: We will only sketch the implementation of function ; the analysis for the function

is identical. Now suppose neither is -closed, 0 1 . We begin by considering an

instance each of MAXCSP 0 and MAXCSP 1 , say 0 and 1 respectively. Both of these

instances are constructed in a manner identical to the instance above. Now we argue that any

solution which satisfies all the constraints in 0 and 1, must set all variables in to 0 and all

variables in to 1.

So we have two functions 0 and 1 such that neither is -closed. Suppose , then

we must have . To see this, consider a satisfying assignment such that 0 ¯ 0;

there must exist such an assignment since 0 is not -closed. Now if , then clearly

at least one constraint corresponding to is unsatisfied - the one in which the positions in are

occupied by the variables in and the positions in are occupied by the variables in

. Thus we must have . But if we have both and ,

then there is at least one unsatisfied constraint in the instance 1 since 1 is not 1-valid. Thus this

case cannot arise.

So we now consider the case . Then for constraints in 0 to be satisfied, we must

once again have ; else there is a constraint with all its inputs set to zero and is hence

unsatisfied. This can now be used to conclude that as follows. Consider a satisfying

assignment with smallest number of ones - this number is positive since 0 is not 0-valid. If we

consider all the constraints corresponding to this assignment with inputs from and only,

it is easy to see that there will be at least one unsatisfied constraint if . Hence each

variable in is set to one in this case. Finally, using the constraints on the function 1 which is

not 1-valid, it is easy to conclude that in fact .


Now let 10 1 be a least hamming weight satisfying assignment for 0; may be zero

but contains at least a single one as 0 is not 0-valid. Then the constraint

0 1 2 1 2

can be satisfied iff 1. Thus all the constraints in 0 and 1 are satisfied along with above

constraint iff 1 and otherwise, we can still satisfy all the constraints in 0 and 1. Hence this

is indeed an implementation of the function . The function can be implemented in an

analogous manner.

4.5.7 REP Helps Implement MAX SNP-Hard Functions

Lemma 21 Suppose is a non-trivial function which is neither 0-valid nor 1-valid. Then

MAXCSP REP implements the XOR function.

Proof: Without loss of generality, assume 0 1 is a satisfying assignment for . We consider

two disjoint set of variables 1 2 and 1 2 . Consider the

MAXCSP REP instance which consists of constraints REP 1 for 2 , constraints

REP 1 for 2 and the constraint 1 2 1 2 . It is now easy to

verify that this yields a 1 -implementation of the function 1 1 if ¯ 1 0 is a satisfying

assignment, and of the function ¯1 1 otherwise. Now an application of the Composition Lemma

yields the lemma.

Corollary 3 Suppose is a non-trivial function which is neither 0-valid nor 1-valid. Then

MAXCSP REP is MAX SNP-hard.

Proof: Immediately follows from Lemma 14 and Lemma 21 above.

4.5.8 Unary Functions Help Implement either REP orMAX SNP-Hard Functions

Lemma 22 Let be a functionwhich is not 2-monotone. ThenMAXCSP can implement

either the XOR or the REP function.


Proof: Since is not 2-monotone and non-trivial, it must be sensitive to at least two variables.

Consider the boolean -cube with each vertex labeled by the function value ; where is the

arity of function . Let denote the set of vertices labeled , 0 1 . If 1 , we claim

that it must be the case that there exists a vertex in which has at least two neighbors in 1 . This

is readily seen using the expansion properties of the -cube; any set of at most 2 1 vertices must

have expansion factor at least one. Furthermore, the expansion factor is precisely one only when

the set induces a boolean 1 -cube. But the later case can’t arise since it would imply that

is a single variable function. Hence there must exist a vertex which has two neighbors in

1 .

Let and be these two neighbors of , differing in the th and the th bit position respectively.

Without loss of generality, we may assume that 1 and 2. Consider now the input instance

which has a constraint of the form 1 2 1 2 2 and constraints of the form

for each appearing in 1 2 and of the form for each appearing in

1 2 . It is now easy to verify that this set of constraints implements one of the

functions 1 2, 1 2, ¯1 ¯2, ¯1 2 or 1 2. The former three implement 1 2

while the later two implement the constraint 1 2.

The following is a straightforward corollary.

Corollary 4 Let be a functionwhich is not 2-monotone. ThenMAXCSP isMAX SNP-

hard.

Proof: If MAXCSP can implement the REP function, then the corollary follows

using the Composition Lemma, the Hardness Lemma and the Lemmas 14 and 16. Otherwise, it

follows from Lemma 15.

Lemma 23 If is a constraint set such that there exist (1) 0 which is not 0-valid, (2) 1

which is not 1-valid and (3) 2 which is not 2-monotone. TheMAXCSP isMAX SNP-hard.

Proof: If either 0 or 1 is -closed then using the Replication Lemma, we can implement the

REP function and using the Composition Lemma along with Lemma 21 allows to conclude that

MAXCSP 0 1 2 implements XOR function.


If neither 0 nor 1 is -closed, then using the Unary Lemma, MAXCSP 0 1 2 can

implement the unary functions and , and then using the Composition Lemma along with

Lemma 22, we conclude that MAXCSP 0 1 2 implements either the XOR function or the

REP function. In the latter case, we can use Lemma 21 to conclude that MAXCSP 0 1 2

can implement the XOR function.

In either of the two situations above, we may conclude by Lemma 14 that MAXCSP 0 1 2

is MAX SNP-hard.

4.6 Hardness at Gap Location 1

It is possible to use a notion closely related to -implementation to conclude from Schaefer’s

dichotomy theorem and show that in the cases where SAT in NP-hard to decide, it is actually

hard to distinguish satisfiable instances from instances which are not satisfiable in a constant fraction

of the constraints. This is termed hardness at gap location 1 by Petrank [89] who highlights the

usefulness of such hardness results in other reductions.

An important characteristic of implementation of a function is that if we are given an assign-

ment to the function variables which does not satisfy , it can always be extended to the auxiliary

variables such that precisely 1 constraints are satisfied. This is a useful feature in establishing

the hardness results for problems such as MAX 2-SAT which do not have hardness gaps located at

1. However, when dealing with problems with hardness gaps located at 1, such as MAX 3-SAT, it

suffices to use a somewhat different notion of -implementations, called weak -implementations1.

A weak -implementation satisfies the condition - of the -implementations and the condition

is replaced by the constraint that the MAXCSP instance implementing it has precisely

constraints. Clearly, weak -implementations can be composed together and they preserve hardness

gaps located at 1.

It is not difficult to verify that Schaefer’s proof is in fact based on weak -implementations of

functions, and hence one may directly conclude from his proof that his class of NP-hard satisfiability

1The name weak -implementation is slightly misleading because this notion is simultaneously both weaker andstricter than the notion of -implementations.


problems are all in fact MAX SNP-hard. This yields Theorem 24.

4.7 Strengthening Schaefer’s Dichotomy Theorem

Schaefer’s proof of NP-hardness in his dichotomy theorem relies on the ability to replicate vari-

ables within a constraint application. We observe that to do so, it suffices to create a weak

implementation of the function REP. Since given a weak implementation, we can replace any

replicated copies of a variable by new variables 1 2 and add constraints of the form

REP 1 2 REP 1 3 REP 1 . We now show how to create a weak implementa-

tion of the REP function.

Now Lemmas 19 and 20 show that MAXCSP 0 1 2 , where 0 is not 0-valid and 1 is

not 1-valid, can be used to create either a weak implementation of the function REP or a weak

implementation of both unary functions and . In the latter case, we can show the following

lemma.

Lemma 24 If is not weakly negative then MAXCSP can weak implement either the

function , or the function . Similarly, if is not weakly positive thenMAXCSP

can weak implement either the function , or the function ¯ ¯.

Proof: We only prove the first part - the second part follows by symmetry. We know that has

a maxterm with at least two positive literals. We consider the function which is existentially

quantified over the variables not in . Let 1 and 2 be the two positive literals in . Set all other

variables in to the value which does not make true. Then the assignment 1 2 0 is a

non-satisfying assignment. The assignments 1 0 2 and 1 0 2 must be satisfying

assignments ny the definition of maxterm. While the assignment 1 2 1 may go either way.

Depending on this we get either the function or .

Corollary 5 If 2 is notweakly positive and 3 is notweakly negative, thenMAXCSP 2 3

weak implements (at gap 1) the XOR function.

Since the SAT problems that we need to establish as NP-hard in Schaefer’s theorem satisfy

the condition that there exists 0 1 2 3 such that 0 is not 0-valid and 1 is not 1-valid,


2 is not weakly positive and 3 is not weakly negative, we conclude that we can weak implement

the XOR function. This, in turn, can be used to create a weak implementation of the function

REP by using the constraints for some auxiliary variable . Thus replication

can be eliminated from Schaefer’s proof.

Chapter 5

The Approximability of Structure

Maximization Problems

103

CHAPTER 5. THE APPROXIMABILITY OF STRUCTURE MAXIMIZATION PROBLEMS104

5.1 Introduction

In this chapter, we continue with our study towards understanding the approximation behavior

of optimization problems. We now consider another class of maximization problems built on

Schaefer’s problems. This new class of optimization problems that we consider, called MAXONES,

captures some newer problems such as MAX CLIQUE, while continuing to capture some of the

old problems like MAX CUT. A problem of this class is also defined by a constraint set . An

instance to this problem consists of constraints on variables. The objective now is to find an

assignment with the largest number of 1’s which satisfies all constraints. A good way to think

about this problem is that we are trying to find the largest subset of a given set of variables among

all “feasible” subsets, i.e., among all subsets that satisfy the given feasibility conditions.

As before, we give a classification theorem to determine the approximability of any problem

in this class from merely its description. Our classification theorem shows that this class contains

problems that are solvable exactly, MAX SNP-complete problems, problems which are approx-

imable to within polynomial factors but no better, and then problems which are not approximable

to within any factor at all. This class thus contains a representative at every known approximability

threshold for the naturally-arising maximization problems1. We feel that this class sheds light on the

approximability of NPO maximization problems in general. Our results here serve to re-affirm the

trends emerging from our study of approximability of these problems. In particular, they provide

some evidence towards the non-existence of natural maximization problems with a log -threshold

for approximability.

The rest of this paper is organized as follows. In Section 5.2, we formalize our problem and

introduce various definitions needed to state our main result. Section 5.3 states the main theorem

and gives an interpretation of our results. Section 5.4 establishes a correspondence between the

approximation of the weighted and the unweighted problems. Finally, in Section 5.5, we describe

the complete proof of our main theorem.

This chapter is based on joint work with Madhu Sudan and David Williamson [71].1Informally speaking, a problem Π has an approximability threshold of if Π is approximable to within a factor

of and is NP-hard to approximate within a factor of .


5.2 Definitions

5.2.1 Problem Definition

We begin by defining our problem of interest; we build on the definitions introduced in the preceding

chapter.

Definition 69 (Weighted MAXONE ) Given a constraint set , the weighted max ones prob-

lem MAXONE is defined as follows:

INPUT: An instance is a triple . It consists of a collection of constraint appli-

cations of the form 1 , on boolean variables 1 , and weights

1 , where , is the arity of , and the ’s are non-negative integers.

OBJECTIVE: Find a boolean assignment 1 satisfying all constraint applications

that maximizes .

If 1 for all , we refer to the problem as the unweighted max ones problem or simply the

max ones problem, and the instance is simply a pair .

We will also sometimes consider the following simpler decision problem.

Definition 70 (SAT ) Given a constraint set , the constraint satisfaction problem SAT is

defined as follows:

INPUT: A collection of constraint applications of the form 1 , on boolean

variables 1 , where , and is the arity of .

OBJECTIVE: Find a boolean assignment 1 satisfying all constraints.

We will be considering the approximability of the MAXONE problems for different con-

straint sets .

Definition 71 For a constraint set and a function : , an -approximation algorithm

for the weighted MAXONE problem is an algorithm which takes as input an instance


of MAXONE with variables, and constraints, and produces an assignment

1 satisfyingall constraints such that for every other feasible solution to the given instance

1 1 . A problem is said to -approximable if it has an -approximation

algorithm with running time bounded by a polynomial in and .

The important feature to notice above is that when is allowed to be a function of the instance

size, then it is measured as a function of only the number of variables . It is important to note that

the number of non-redundant constraints in any MAXONE problem is bounded by a polynomial in

the number of variables. Hence fixing to be a function of alone does not result in much loss of

information.

Before we can state our precise result, we need some characterizations of constraint functions.

Thus we begin with some definitions needed to describe our main result.

5.2.2 Constraint Functions

We need the following additional definitions of contraint functions.

Definition 72 (Affine with Width 2) A function is affine with width 2 if it can be expressed as a

conjunction of linear equalities over GF(2) with at most two variables per equality constraint.

Definition 73 (2CNF) A function is a 2CNF function if it can be expressed in conjunctive normal

form with all disjuncts having at most two literals.

We add one more definition to the above collection.

Definition 74 (Strongly 0-Valid) We say a function is strongly 0-valid if it is satisfied by any

assignment with less than or equal to 1 ones.

5.2.3 Families of Constraint Functions

We use the following shorthand notation:

Let 1 be the family of all 1-valid functions.


Let 2 be the family of all weakly positive functions.

Let 3 be the family of all affine functions of width 2.

Let 4 be the family of all affine functions.

Let 5 be the family of all strongly 0-valid functions.

Let 6 be the family of all weakly negative functions.

Let 7 be the family of all 2CNF functions.

Let 8 be the family of all 0-valid functions.

5.3 Overview of the Main Result

In this section, we formally state and interpret our results.

5.3.1 Main Theorem

Our main theorem is as follows.

Theorem 25 1. If for some 1 2 3 then the weighted MAXONE is in P.

2. If 4 and for every 1 2 3 then the weighted MAXONE is

APX-complete.

3. If for some 5 6 7 and for every 1 2 3 4 then weighted

MAXONE is poly-APX complete.

4. If 8 and for every 1 7 , then is decidable but finding a

solution of positive value is NP-hard.

5. If is not a subset of for any 1 8 , then finding any solution satisfying a given

set of constraints is NP-hard.

For contrast, we restate Schaefer’s result here.


Theorem 26 (Schaefer [95]) If for some 1 2 4 6 7 8 , then SAT is in P.

Otherwise SAT is NP-hard.

Notice that Case 5 of our main theorem follows immediately from Schaefer’s theorem.

5.3.2 Discussion

A useful interpretation of the problems studied here is as follows. We study the approximation

properties of integer programming problems, say IP( ), such that the objective function is of the

form MAX , ’s are the weights associated with the variables, ’s are 0/1 variables and the

constraints are boolean functions of bounded arity drawn from the given family . Our main result

shows that the precise asymptotic approximability of each problem in this class can be identified

merely from the syntactic structure of the constraint set . We have already seen that certain integer

programming frameworks when combined with restriction on the input structure, yield polynomial

time approximation schemes. Barland, Kolaitis and Thakur [14] recently initiated a study of

syntactic restrictions on a class of integer programs which enable constant factor approximations.

An interesting aspect of our results is that the approximability of the problems in this class is

essentially independent of the weights. That is, any hardness of approximation result holds for the

unweighted instances arising in IP( ), while the approximability results apply to arbitrary weighted

instances arising in IP( ).

5.4 Preliminaries

In this section, we prove a few preliminary lemmas that we will need in the proof of our main

theorem, particularly in Cases 2 and 3. We first show that in these cases, it is essentially equivalent

for us to consider the weighted or unweighted MAXONE problem. We then show that our ability

to work with the weighted MAXONE allows us in these cases to use a notion of “implementing”

a constraint not in in terms of other constraints in . The idea of implementing constraints is

adapted from Chapter 4, and is used heavily in our proofs of Cases 2 and 3.


5.4.1 Weighted vs. unweighted problems

We begin with a slightly stronger definition of polynomial-time solvability of SAT that we

will need. We then show that given this stronger form of SAT that insofar as APX-hardness

and poly-APX-hardness are concerned, the weighted and unweighted cases of MAXONE are

equivalent. We conclude by showing that in Cases 2 and 3 the stronger form of SAT holds.

Definition 75 We say that a constraint satisfaction problem SAT is strongly decidable if given

constraints on variables 1 and an index 1 , there exists a polynomial time

algorithmwhich decides if there exists an assignment to 1 satisfying all constraints and

additionally satisfying the property 1.

Lemma 25 For every strongly decidable function family , for every of the form 1 for some

positive integer and for every non-decreasing function : , -approximating the

weighted MAXONE problem reduces to -approximating the (unweighted) MAXONE

problem, where 1 .

Proof: Given an instance of a weighted MAXONE problem, we create a sequence

of instances of weighted MAXONE problems with the final instance having 2 variables,

with the weight of every variable being 1, and with the property that given an -approximate

solution to the final instance we can obtain an -approximate solution to the instance .

Assume w.l.o.g. that 1 2 . From here onwards will denote the smallest

index such that there exists a feasible solution to with 1. Notice that can be computed in

polynomial time. The instance 1 is defined to be , where if and 0

otherwise. Since no solution of has 1 for , this change in the weight function does not

alter the value of any solution. Thus 1 is essentially equivalent to . Notice that the weight of the

optimal solution to 1 is at least .

The instance 2 is defined to be , where is rounded up to the nearest integral

positive multiple of . Notice that the net increase to the weight of any solution by this increase

in the weights is at most .


Finally the (unweighted) instance 3 has as its variables copies of every variable , where

. Notice that by the definition of ’s, the ’s are integral and 1 . Given

a constraint on variables 1 , the instance 3 has 1 2 copies of the

constraint applied to all different collections of -tuples of variables containing a copy of the

1th variable in the first position, 2th variable in the second position and so on. The weight of all

variables is 1.

Let 1 denote the number of variables in 3. We will now show that given -

approximate solution to 3, we can reconstruct a -approximate solution to . It is easy to verify

that any solution to 3 can be modified to have all copies of a variable assigned to 1 or all assigned

to zero, without changing any 1 to a 0. Thus every solution to 2 of weight can be transformed to

a solution of weight for 3 and vice versa. Thus given a solution of weight at least OPT 3 ,

we can reconstruct (in polynomial time) a solution of weight at least OPT 2 for 2. To conclude

the argument it suffices to show that this can be used to construct an -approximate solution to

1.

Our candidate solutions will be the solution which assigns 1 to and the solution to the instance

2. If the value of the -approximate solution to 3 is , then the value of a solution so returned

is at least max which is at least 1 . (The inequality max 1 follows

from a simple averaging argument.) Since OPT 2 and OPT 1 OPT 2 , we find

that this solution has value at least OPT 11

OPT1 . Thus this solution is a 1

approximate solution to .

Finally observe that since 2 , this is also an 1 2 -approximate solution

to the instance .

As our examination will eventually show, there is really no essential difference in the approx-

imability of the weighted and unweighted problems. For now we will satisfy ourselves by stating

this conditionally.

Corollary 6 For any strongly decidable function family , theMAXONE problem is APX-hard

if and only if the weighted MAXONE problem is APX-hard.


Corollary 7 For any strongly decidable function family , the MAXONE problem is poly-

APX-hard if and only if the weighted MAXONE problem is poly-APX-hard.

Before concluding we show that most problems of interest to us will be able to use the equiva-

lence between weighted and unweighted problems.

Lemma 26 If for some 1 7 , then is strongly decidable.

Proof: For a function and an index 1 , let 1 be the function:

1 1def

1 1 1 1

Further let be the family:

def1

First observe that the problem described in the lemma can be expressed as a problem of SAT .

Further, observe that if for 1 2 3 4 6 7 , then as well. Lastly, if 5,

then 8. Thus in each case we end up with a problem from SAT which is in by

Schaefer’s theorem.

5.4.2 Implementations and weighted problems

The ability to work with weighted problems now allows us to use existential quantification over

auxiliary variables and the notion of functions “implementing” other functions. We start with the

definition of implementing a constraint — an adaptation of the definition given in the preceding

chapter.

Definition 76 [Implementation]A set of constraint applications 1 using constraints from

over a set of variables 1 2 and 1 2 form an implementation

of a boolean function iff the following conditions are satisfied:

(a) for any assignment of values to such that is true, there exists an assignment of values

to such that all the constraint applications are satisfied, and


(b) for any assignment of values to such that is false, no assignment of values to can

satisfy all the constraints.

A constraint set implements a constraint if there exist a implementation of using constraints

from .

The following lemma establishes the utility of implementations in showing hardness of approx-

imating MAXONE problems.

Lemma 27 Given function families , such that the weighted MAXONE problem has a

-approximation algorithmand every function can be implemented by the family , then

there exist constants such that theweightedMAXONE problemhas a -approximation

algorithm.

Proof: Let and be the maximum arity of any function . Let be the largest

number of auxiliary variables used in implementing any function by . Notice is a finite

constant for any fixed , .

Given an instance with constraints on variables of MAXONE , we create

an instance of MAXONE as follows: has the variables 1 of and in addition

“auxiliary” variables 1 1. The weights corresponding to 1 is 1 (same as

in ) and the auxiliary variables have weight zero. The constraints of implement the constraints

of . In particular the constraint 1 of is implemented by a collection of constraints

from (as dictated by the implementation of by ) on the variables 11 .

By the definition of an implementation, it is clear that the every feasible solution to can be

extended (by some assignment to the variables) into a solution to . Alternately every solution

to immediately projects into a solution of . Furthermore, the value of the objective function

is exactly the same (by our choice of weights). Thus a -approximate solution to gives a

-approximate solution to .

It remains to study this approximation as a function of the instance size. Observe that the

instance size of is much larger. Let denote the number of variables in . Then is

upper bounded by , where is the number of constraints in . But , in turn, is at


most . Thus 1 , implying that an -approximate solution to , gives an

1 -approximate solution to . Thus an -approximation algorithm for the weighted

MAXONE problem yields an -approximation for the weighted MAXONE -problem,

for 1 and .

The main corollaries of the above lemma are listed below.

Corollary 8 If weighted MAXONE is APX-hard and implements every function in , then

weighted MAXONE is APX-hard.

Corollary 9 If weighted MAXONE is poly APX-hard and implements every function in ,

then weighted MAXONE is poly APX-hard.

Lastly we describe one more tool that comes in useful in creating reductions. This is the notion

of implementing a property which falls short of being an implementation of an actual function.

The target functions in the following definitions are the functions which force variables to being

constants (either 0 or 1). However, sometimes we are unable to achieve this. So we end up

implementing a weaker form which however suffices for our applications. We next describe this

property.

Definition 77 [Existential Zero] A family of constraints can implement the existential zero

property if there exists a set of constraints 1 over variables and an index

1 such that the following hold:

There exists an assignment 1 to 1 such that assigning 0 to the first

variables 1 satisfies all constraints.

Conversely, every assignment satisfying all the constraints must make at least one of the

variables in 1 zero.

An Existential One can be defined similarly.

Definition 78 Given a constraint of arity and a set 1 , the constraint 0 is a

constraint of arity given by 0 1 1 0 0 2 0 0 , where


the zeroes occur in the indices contained in . For a constraint set , the 0-closure of , denoted

0 is the set of constraints 0 1 .

Notice that 0 essentially implements every function that can be implemented by 0 ,

except the function 0 . We define 1 similarly. Then 0 1 0 1.

Lemma 28 If a family of constraints can implement the existential zero property, then can

implement every function in the family 0.

Proof: We show how to implement the function 0 1 1 . The proof can be ex-

tended to other sets by induction. Let the functions 1 implement the existential zero

property on variables 1 with auxiliary variables 1 . Then the constraints

1 1 , for 1 , along with the constraints 1 on the variables

1 implement the constraint 0 1 1 . (Observe that since at least one of the ’s

in the set 1 is zero, the constraint 0 1 1 is being enforced.) Furthermore, we

can always set all of 1 to zero, ensuring that any assignment to 1 1 satisfying

0 1 1 does satisfy all the constraints listed above.

Using the same proof as above, we also get the following lemma.

Lemma 29 If a family of constraints can implement the existential one property, then can

implement every function in the family 1.

5.5 Proof of Main Theorem

We start with the sub-cases that are easier to prove and then move on to the more difficult sub-cases.

We first tackle the cases 1, 4 and 5 of Theorem 25. Recall that we already showed that Case 5

follows from Schaefer’s theorem.

5.5.1 Case 1: The Polynomial-Time Case

Lemma 30 The weighted MAXONE problem is in P if each is 1-valid.


Proof: Set each variable to one; this satisfies all constraints with a maximum possible weight

of solution.

Lemma 31 The weighted MAXONE problem is in P if each is weakly positive.

Proof: Consider the CNF formulae for the such that each clause has at most one negated

variable. Clearly, clauses consisting of a single literal force the assignment of these variables.

Setting these variables may create new clauses of a single literal; set these variables and continue

the process until all clauses have at least two literals or until a contradiction is reached, in which

case no feasible assignment is possible. In the former case, setting the remaining variables to one

satisfies all constraints, and there exists no feasible assignment with a greater weight of ones.

Lemma 32 The weighted MAXONE problem is in P if each is an affine function of

width 2.

Proof: We reduce the problem of finding a feasible solution to checking whether a graph is

bipartite, and then use the bipartition to find the optimal solution. Notice that each constraint

corresponds to a conjunction of constraints of the form or . Create a vertex for

each variable and for each constraint , add an edge . For each constraint ,

identify the vertices and ; if this creates a self-loop, then clearly no feasible assignment is

possible. Check whether the graph is bipartite; if not, then there is no feasible assignment. If so,

then for each connected component of the graph choose the larger weight side of the bipartition,

and set the corresponding variables to one.

5.5.2 Case 4: The Decidable but Non-Approximable Case

Lemma 33 If 8, then SAT is in P.

Proof: Follows trivially since every instance is satisfiable. The assignment of 0’s to every

variable is a satisfying assignment to every instance.

Lemma 34 If , for any 1 7 , then the problem of finding solutions of non-zero

value to a given instance of (unweighted) MAXONE is NP-hard.


Proof: Given a function : 0 1 0 1 and an index , let be the function

mapping 0 1 1 to 0 1 given by

1def

1 1 1 1 1 1 0 1

Let be the set of functions defined as follows:

def

I.e., is the set of functions obtained by setting none or one of the variables of to 1. We will

argue that deciding SAT is NP-hard and then that deciding SAT reduces to finding non-zero

solutions to MAXONE .

First observe that , for 1 8 . In particular it is not 0-valid, since is not

strongly 0-valid. Hence, once again applying Schaefer’s result, we find that deciding SAT is

NP-hard.

Now given an instance of SAT on variables 1 with constraints 1 , with

1 and 1 , consider the instance of MAXONE defined

on variable set

1 1 1 1

with the following constraints:

1. Let be a non-1-valid function in . We introduce the constraint 1 .

2. For every constraint 1 , 1 , we introduce two constraints 1

and 1 .

3. For every constraint 1 1 , 1 , we introduce 2 1 constraints.

For simplicity of notation, let 1 1 1 1 1 0 1 1

where . The 2 1 constraints are:

1 1 , for 1 1.

1 1 , for 1 .


1 1 , for 1 1.

1 1 , for 1 .

We now show that the instance of MAXONE created above has a non-zero satisfying

assignment if and only if the instance of SAT has a satisfying assignment. Let 1 2

be a satisfying assignment for the non 1-valid function chosen above. First if 1 form

a satisfying assignment to the instance of SAT , then we claim that the assignment

for 1 , 1 1 and for 1 is a satisfying assignment to the

instance of MAXONE which has at least one 1 (namely 1). Conversely, let some non-

zero setting 1 1 1 1 satisfy the instance of MAXONE . W.l.o.g./

assume that one of the variable 1 1 1 is a 1. Then we claim that the setting

, 1 satisfies the instance of SAT . It is easy to see that the constraints

1 , 1 , are satisfied. Now consider a constraint 1 1

0 1 1 1 1 1 . Since at least one of the variables in the set 1

is a 0 and at least one of the variables in the set 1 1 1 is 1, we know that both

0 1 1 and 1 1 1 are satisfied and hence 1 1 1. Thus the

reduced instance of MAXONE has a non-zero satisfying assignment if and only if the instance

of SAT is satisfiable.

5.5.3 Case 5: The NP-Hard Case

Lemma 35 If, for all 1 8 , , then deciding SAT is NP-hard.

Recall that we have already indicated that this follows from Schaefer’s result [95].

5.5.4 Case 2: The APX-Complete Case

We now continue with the proof of Theorem 25; we first turn towards the proof of Case 2.

5.5.4.1 Membership in APX

Lemma 36 If 4, then the weighted MAXONE problem is in .


Proof: By Lemmas 26 and 25 it suffices to consider the unweighted case. In this case when

all constraints are affine, then satisfying all constraints is essentially the problem of solving a linear

system of equations over GF[2]. If the system is overdetermined, then no feasible solution exists.

If the system is exactly determined, then the setting of all variables is forced, and we find the

assignment with the maximum possible number of ones. If the system is underdetermined, then

setting some number of variables arbitrarily determines the remainder of the solution; to be more

precise, the variables can be partitioned into and such that for some 0/1 matrix

and some 0/1 vector (where matrix arithmetic is carried out over GF[2]). Setting the variables

in to 1 with probability 1/2 thus ensures that the probability of each variable is 1 with probability

1/2. The expected number of ones is 2, no worse than a factor of two from the maximum number.

5.5.4.2 APX-Hardness

Definition 79 [ -closed] A constraint is called -closed (or complementation-closed) if for all

assignments , ¯ .

Definition 80 The constraint XOR is defined to be the constraint XOR 1 if and only if

1. The constraintREP is defined to be the constraintREP 1 if and only if 0.

Lemma 37 Given a constraint which is not 1-valid, the following hold:

1. If is -closed, then implements REP and XOR.

2. If is a constraint which is not -closed, then implements an existential zero and an

existential one.

3. either implements an existential zero or implements the functions REP and XOR.

Proof: If is -closed, then it is not 0-valid. In this case, we can use the Replication Lemma

from Chapter 4 to implement REP and XOR. The general idea there is to create two sets of variables

and . For each satisfying assignment of with ones and zeroes, create all possible

constraints by applying to all possible sets of variables such that if 0 then the th variable is a


variable from , otherwise it is a variable from . One can then show that all variables in must

all be set to the same value, and all variables in to the opposite value. The constraint REP can be

enforced by considering variables from the same set, and XOR by considering variables from both

and . This gives us Part (1) above.

For Part (2), if is 0-valid, then the constraint 1 implements an existential zero.

Hence we can assume that is neither 0-valid nor 1-valid. We follow the proof of the Unary Lemma

from Chapter 4. The general idea there creates two sets of variables and as above. We now

use the non C-closed function to impose the constraint that the variables in must be set to zero,

and those in to one, and hence implement the 1 and 0 functions. In particular we can impose the

0 and 1 functions, which gives the existential zero and one properties.

Part (3) follows from the fact that if is C-closed, then Part (1) applies, else we can apply Part

(2) to .

Definition 81 The function XOR3 is defined to be the function XOR3 1 if and only if

0. The function XOR4 is defined to be the function XOR4 1 if and only

if 0.

Lemma 38 The weighted problem MAXONE XOR3 is APX-hard.

Proof: We reduce the MAX CUT problem to the weighted MAXONE XOR3 problem as

follows. Given a graph we create a variable for every vertex and a variable

for every edge . The weight associated with the vertex variable is 0. The weight of

an edge variable is 1. For every edge between and we create the constraint 0.

It is clear that any 0 1 assignment to the ’s define a cut and for an edge , is one iff

and are on opposite sides of the cut. Thus solutions to the MAXONE problem correspond to

cuts in with the objective function being the number of edges crossing the cut.

Lemma 39 The weighted problem MAXONE XOR4 XOR is APX-hard.

Proof: The proof is similar to that of Lemma 38. In this case, given a graph , we

create the variables for every , for every and one global variable (which


is supposed to be zero) and auxiliary variables 1 . For every edge

in we impose the constraints 0. In addition we throw in the constraints

1 for every 1 . Finally we make the weight of the vertex variables and

zero and the weight of the edge variables and the auxiliary variables is made 1. The optimum

to the so created MAXONE problem is MAX CUT . Since MAX CUT 2, this

preserves APX-hardness.

Lemma 40 Suppose implements the existential zero property and is the constraint 1

for some integer 3 and some 0 1 . Then the family implements

XOR3.

Proof: Since implements the existential zero property, the set can implement

0 (using Lemma 28). In particular, can implement the constraints 1 2 and

1 2 3 . Notice finally that the constraints 1 2 and 3 implement

the constraint 1 2 3 0. Thus implement the constraint XOR3.

Lemma 41 Suppose implements the XOR function and is the constraint 1

for some integer 3 and some 0 1 . Then the family either implements XOR3 or

XOR4.

Proof: First observe that can also implement the function REP since 1 1 and 2 1

implement the constraint 1 2 0.

Next observe that we can assume w.l.o.g. that 0. If not the constraints 1 1 1

and 1 implement the function 1 0.

Now if is odd, then the constraints 1 0 and REP 4 5 , REP 6 7 and so

on up to REP 1 implement the constraint 1 2 3 0.

Lastly, if is even, then the constraints 1 0 and REP 5 6 , REP 7 8 and

so on up to REP 1 implement the constraint 1 2 3 4 0.

Theorem 27 If 4 and for every 1 2 3 then MAXONE is APX-hard.

Proof:[Sketch] Let 1 be a non width-2 affine function and let 1 be a

linear constraint of size greater than 2. Then the constraint 1 s.t. 1 1


implements the constraint 1 2 0 1. Now let be a non-1-valid function.

If implements the existential zero property, then by Lemma 40 the set can implement

XOR3. Else, by Lemmas 37 and 41, implements XOR and either implements XOR3 or

XOR4 . Thus in any case can either implement the function XOR3 or the set XOR XOR4 ,

either of which is APX-hard.

5.5.5 Case 3: The Poly-APX-Complete Case

5.5.5.1 Membership in poly-APX

Lemma 42 If for some 1 2 3 4 5 6 7 then MAXONE can be approximated to

within a factor of .

Proof: Schaefer’s results imply a polynomial time algorithm to compute a feasible solution.

If the feasible solution has at least one 1, we are done. Else, iteratively try setting every variable

to one and computing a feasible solution. Note that if is affine (or 2CNF), then the functions

obtained by restricting some variable to be 1 remains affine (or resp. 2CNF), and thus this new

class is still decidable. Lastly, a strongly 0-valid family remains 0-valid after this restriction and is

still decidable. If the decision procedure gives no non-zero solution, then the optimum is zero, else

we output a solution of value at least 1.

5.5.5.2 poly-APX-hardness

The lemma below shows that there exists a non-C-closed function in . Since there also exists a

non-1-valid function in , by the proof of Lemma 37, we can implement an existential zero.

Lemma 43 If for some 5 6 7 , but for any 1 2 3 4 , then there exists

a non-C-closed function in .

Proof: Notice that a C-closed 0-valid function is also 1-valid. A C-closed weakly positive

function will also be weakly negative. Lastly a C-closed 2CNF function is a affine function of

width 2. Thus if is 0-valid, then the function which is not 1-valid is not C-closed. If is weakly


negative, the function which is not weakly positive is not C-closed. Similarly if is 2CNF, then

the function that is not affine is not C-closed.

As a consequence we obtain the following lemma.

Lemma 44 If for some 5 6 7 , but for any 1 2 3 4 , then either

implements every function in 0 1 or implements 0 and every function in is 0-valid.

Proof: By Lemma 43 there exists a non C-closed function in . Also since is not 1-valid,

there exists a function which is not 1-valid. By Lemma 37 the set can implement

an existential zero and hence can implement every function in the family 0. Furthermore, if

contains a non 0-valid function then it can implement 0 1 (again by Lemma 37), else every

element of is 0-valid.

Our goal will be to implement the function ¯ 1 ¯ , for some 2. The following

lemma shows that this will imply poly-APX-hardness.

Lemma 45 If ¯1 ¯ , then MAXONE is poly-APX-hard.

Proof: We do a reduction from clique. Given a graph , construct a MAXONE instance

consisting of a variable for every vertex in and the constraint function is applied to every subset

of vertices in which does not induce a clique. It may be verified that the optimum number

of ones in any satisfying assignment to the instance created in this manner is max 1 ,

where is the size of the largest clique in . Given a solution to the MAXONE instance

with ones, it is easy to obtain a clique in with 1 vertices. Thus the existence of an

-approximation algorithm for the MAXONE problem implies the existence of a 1 -

approximation algorithm to the clique problem. The poly-APX-hardness of clique now implies the

lemma.

We need to consider two cases: the case in which there exists a function that is not 0-valid, and

the case in which all functions are 0-valid. In the former case, due to Lemma 44, we can afford

to work with the family 0 1. We first show that in this case, we can implement ¯ ¯. We will

then turn to the case in which all functions are 0-valid and show that we can either implement

¯1 ¯ or an existential one, and this will complete the proof of poly-APX-hardness.


Recall that we have a function in that is not weakly positive and a function that is not affine.

Case A : MAXONE implements an existential one. Lemmas 46-50 cover this case.

Lemma 46 If is not weakly positive, then the family 0 1 implements either XOR or ¯ ¯.

Proof: Let ¯1 ¯ 1 be a maxterm in with more than one

negation i.e. 2. Substituting a 1 in place of variables ¯3 ¯4 ¯ , a 0 in place of variables

1 2 , and existentially quantifying over all variables not in , we get a function such

that ¯1 ¯2 is a maxterm in .

By definition of maxterm, must be satisfied whenever 1 2 1. Now if is also satisfied

when 1 2 0, we get the function ¯1 ¯2, else we get the function XOR 1 2 .

Lemma 47 The function (family) XOR can implement the function REP.

Proof: To implement REP , we include the constraints XOR and XOR for an

auxiliary variable .

Lemma 48 Let be a non-affine function. Then the function family REP XOR 0 1 can either

implement the function ¯ or ¯ ¯ .

Proof: Since is non-affine, we essentially have the following situation for three satisfying

assignments 1 2 and 3 for .

1 00 0 00 0 00 0 00 0 11 1 11 1 11 1 11 1 1

2 00 0 00 0 11 1 11 1 00 0 00 0 11 1 11 1 1

3 00 0 11 1 00 0 11 1 00 0 11 1 00 0 11 1 1

1 2 3 00 0 11 1 11 1 00 0 11 1 00 0 00 0 11 1 0

00 0 ¯ ¯ ¯ ¯ ¯ ¯ ¯ ¯ ¯ 11 1

Fixing the above variables to 0’s and 1’s as shown in the last row, and assigning replicated

copies of three variables and (and their negations using XOR), we get a function

with the following truth-table :


xyz

00 01 11 10

1 B 1 A0

1 C 1 D 0

Figure 5.1: Truth-table of the function

The undetermined values in the table are indicated by the parameters and . The

following analysis shows that for every possible value of these parameters, we can indeed implement

an OR function using the constants 0 and 1.

0 ¯

1 0 0 ¯

1 1 0 0 ¯

1 1 1 0 1 ¯ ¯

1 1 1 1 1 ¯

Lemma 49 The family ¯ XOR implements the function ¯ ¯.

Proof: To implement ¯ ¯, we create an auxiliary variable . We now add two constraints,

namely ¯, and XOR . Clearly, all constraints are satisfied only if ¯ ¯ is satisfied.

We now summarize the effect of Lemmas 46-49 formally.

Lemma 50 If for any 1 2 3 4 , then 0 1 implements the function ¯ ¯.

Proof: Lemma 46 yields that 0 1 can either implement ¯ ¯, in which case we are done, or

it can implement XOR. By Lemma 47 this implies 0 1 can also implement REP. Lemma 48 and

non-affineness of imply in turn that in this case, 0 1 can either implement ¯ ¯ or it implements

¯. In the former case we are done, and in the latter we use Lemma 49 and the fact that 0 1

can implement both XOR and ¯ to conclude that it can implement ¯ ¯.


We now turn to the case in which all functions are 0-valid, and show that either we can implement

¯ ¯, or implement a 1. If the former occurs, we are done, and if the latter, we can apply Lemma 50.

Case B : All functions are 0-valid.

Lemma 51 If is 0-valid and not weakly positive, then the family 0 either implements ¯ 1

¯ for some 2 or it implements ¯ or REP.

Proof: Let ¯1 ¯ 1 be a maxterm in with more than one negation

i.e. 2 (such a maxterm exists since is not weakly positive). Substituting a 0 in place of

variables 1 2 , and existentially quantifying over all variables not in , we get a function

such that ¯1 ¯2 ¯ is a maxterm in . Consider an unsatisfying assignment for

with the smallest number of 1’s and let denote the number of 1’s in ; we know 0 since the

original function is 0-valid. WLOG assume that assigns value 1 to the variables 1 2

and 0’s to the remaining variables. It is easy to see that by fixing the variables 1 2

to 0, we get a function ¯1 ¯2 ¯ . If 1, then this implements the function

¯1 ¯ and we are done.

Otherwise 1, i.e. there exists an unsatisfying assignment which assigns value 1 to exactly

one of the ’s, say 1. Now consider a satisfying assignment which assigns 1 to 1 and has a

minimum number of 1’s among all assignments which assign 1 to 1. The existence of such an

assignment easily follows from being a maxterm in . W.l.o.g. assume that 1 0 . Thus

the function looks as follows:

1 2 3 1

1 0 0 00 0 00 0 1

2 1 0 00 0 00 0 0

3 1 1 11 1 00 0 1

4 0 1 00 0 ?

Existential quantification over the variables 3 4 and fixing the variables 1 through

to 0 yields a function which is either 1 ¯2 or REP 1 2 . The lemma follows.


If we can implement ¯, then the following lemma shows that we can essentially implement

a 1, and thus we can reduce to the previous case.

Lemma 52 If MAXONE ¯ is -approximable for some function , then so is

MAXONE 1 .

Proof: Given an instance of MAXONE 1 we construct an instance of MAXONE

¯ as follows: The variable set of is the same as that of . Every constraint from in

is also included in . The only remaining constraints are of the form 1 for some variables

(imposed by the function 1). We simulate this constraint in with 1 constraints of the form

¯ for every 1 , . Every non-zero solution to the resulting instance is

also a solution to , since the solution must have 1 or else every 0. Thus the resulting

instance of MAXONE ¯ has the same objective function and the same feasible space

and is hence at least as hard as the original problem.

Now by Lemma 51 the only remaining subcase is if we can implement REP. The following two

lemmas show that in this case we can either implement ¯ ¯ or ¯. If we can do the former, we

are done, and if the latter, we can use ¯ to implement a 1, and reduce to the previous case.

Lemma 53 Let be a non-affine function. Then there exist two satisfying assignments 1 and 2

such that 1 2 is not a satisfying assignment for .

Proof: See, for example, Schaefer [95].

Lemma 54 If is 0-valid function and non-affine, then MAXONE REP implements either

the function ¯ ¯ or the function ¯ .

Proof: Using Lemma 53 and the fact that is 0-valid, we essentially have the following

situation:


00 0 00 0 00 0 00 0 1

1 00 0 00 0 11 1 11 1 1

2 00 0 11 1 00 0 11 1 1

1 2 00 0 11 1 11 1 00 0 0

00 0

Fixing the above variables to 0’s as shown in the last row, and assigning replicated copies of

three variables and , we get a function with the following truth-table :

xyz

00 01 11 10

1 B 1 A0

1 C 1 D 0

Figure 5.2: Truth-table of the function

The lemma now follows using an analysis identical to the one used in Lemma 48.

We are now ready to prove the final theorem for this section.

Theorem 28 If for any 1 2 3 4 , then MAXONE is poly-APX-hard.

Proof: As usual it suffices to consider the weighted problem. We will show that either

implements ¯1 ¯ for some 2, or that implements ¯ and 1 implements ¯ ¯.

The theorem follows from an application of Lemma 52, which shows that MAXONE is as hard

to approximate as MAXONE 1 , which in turn is poly-APX hard.

By Lemma 44 we have that implements 0. Furthermore, if is not 0-valid, then

implements 0 1 which in turn implements ¯ ¯ (by Lemma 50). Hence we are left with the case

where is 0-valid. In this case, by Lemma 51, 0 (and hence ) implements ¯1 ¯ or

¯ or REP. In the first case we are done. In the second case we get the second possibility from

above since 1 implements 0 1 and 0 1 implements ¯ ¯ (from Lemma 50 again). Finally if


0 can implement REP, Lemma 54 can be applied to claim that either implements ¯ ¯ or 0

implements ¯. If implements ¯ ¯ then we are done, else it implements ¯. Again we

apply Lemma 50 to conclude that in this case also implements ¯ and 1 implements ¯ ¯.

Chapter 6

The Approximability of Graph Coloring

129

CHAPTER 6. THE APPROXIMABILITY OF GRAPH COLORING 130

6.1 Introduction

Given a graph , a legal coloring of is an assignment of a color to each vertex of , such that no

edge of is connecting two vertices that are assigned the same color. The chromatic-number of ,

denoted is the minimum number of colors necessary for a legal coloring of .

Given as input a graph , one may want to come up with a legal coloring of with the fewest

possible colors. A polynomial-time algorithm that achieves coloring with a set of colors whose

size is not vastly larger than the optimum ( ) has many practical applications. As an example

consider the following computational problem: given a set of tasks to perform, where some of the

tasks are pair-wise conflicting (say they cannot be carried out at the same time or at the same place),

find a partition of the set of tasks such that no set of the partition contains two conflicting tasks.

This problem is equivalent to the chromatic number problem: consider the graph whose vertices are

all tasks, where two vertices are connected if the corresponding tasks are conflicting. A coloring of

this graph is a non-conflicting partition of the set of tasks.

Coloring a graph with the minimum colors was shown to be NP-hard by Karp [63]; the

results there imply that coloring a 3-colorable graph with 3 colors is NP-hard (this implies the same

hardness result for -colorable graph for any 3).

On the other hand, Blum [19, 20], following earlier work by Wigderson [98] provided a

polynomial-time algorithm which, on input a 3-colorable -vertex graph, finds a legal coloring

using at most 3 8 log85 colors. More recently, Karger, Motwani and Sudan, designed a

randomized polynomial time algorithm which uses no more than 1 4 log . Their techniques

also apply to -colorable graphs where the guarantee is 1 31 colors. In general, the best

known performance guarantee is due to an algorithm by Halldorsson [49] that colors a graph withlog log 2

log3 colors.

These results leave a huge gap between the achievable coloring and the apparently intractable.

In this chapter, we slightly narrow this gap by strengthening Karp’s hardness result. In particular, we

show that separating between the case where 3 and the case where 5 is NP-hard.

This implies that given a 3-colorable graph , it is NP-hard to color with 4 colors. An immediate

This chapter is based on joint work with Nati Linial and Muli Safra [66].


corollary of this result is that for any fixed 3, it is NP-Hard to color a -colorable graph with

2 3 1 colors.

Some results on the hardness of approximate coloring were obtained over the years. Garey and

Johnson [40] have used composition operations on graphs to show that if, for every , there exists a

polynomial time procedure to color a -colorable graph with fewer than 2 6 colors then P=NP.

However, they do not specify an integer for which it is NP-hard to color a -colorable graph

with fewer than 2 6 colors. Furthermore, their result clearly says nothing about the hardness of

coloring a 3-colorable graph. Linial and Vazirani [78] have used squares of graphs to investigate

the best ratio of approximating the chromatic number in polynomial time. They give evidence that

this ratio, for -vertex graphs is either below log1 1 or above Ω 1 .

Building on the hardness of approximation results discovered through the PCP-based character-

ization of NP, Lund and Yannakakis [82] showed that for certain two constants 0 1 2 1, it

is NP-hard to color an 1 -colorable graph with 2 colors. The authors go on to prove that for every

constant there exists a constant such that it is NP-hard to color a -colorable graph by

colors. However, depends on and is relatively large in comparison; this is again not applicable

to small fixed values of the chromatic number. More recently, Feige and Kilian [38] have shown a1 -hardness for the general graph coloring problem.

The remainder of this chapter is organized as follows. In Sections 6.2 and 6.3, we present

a simple proof to both the theorems of Lund and Yannakakis [82] concerning the hardness of

approximating the chromatic number. In Section 6.4 we prove that, given a graph , it is NP-Hard

to distinguish between the case where 3 and the case where 5, and therefore, it is

NP-hard to color a 3-colorable graph using only 4 colors.

6.2 Hardness of Approximating the Chromatic Number when the Value

May be Large

In this section, we give a simple proof that chromatic number is hard to approximates to within a

factor of , while the value of the chromatic number itself may be for some constant 0 1.


Notation

We use to denote the size of the largest clique in a graph . The chromatic number of

is denoted by , while its clique-cover number is denoted by ¯ . We use ¯ to denote the

complement graph of . Since any clique-cover of a graph (a partition of into cliques) is a

legal coloring of the graph ¯ (a partition of ¯ into independents sets), it follows that ¯ ¯ .

The Starting Point

We start with an -partite graph that results from the reduction of [37], so it either has an -clique

(the maximum possible) or every clique of is of small size.

Pictorially, one can think of the graph as consisting of rows, each being an independent

set, and having some edges connecting vertices in different rows. Hence a clique of size in

has exactly one representative in each row. Note that each row of may have different number of

vertices.

Let be the maximum number of vertices in any row of , and assume that either or

for some absolute constant , 0 1. For any , separating between these

two cases is NP-hard [37, 8, 7, 90].

Our Approach

Given such a graph with for some constant , 0 1, we construct a graph with

vertices for some 1 such that ¯ when and ¯ otherwise.

Since ¯ ¯ , our result follows from the construction of such a graph .

The graph can be intuitively described as a transformation of which makes symmetric

under certain rotations, so that an -clique has several symmetric images which together cover

completely. On the other hand, the transformation also ensures that the size of the largest clique in

is at most that of .

The graph , like , is an -partite graph, and is constructed so that for any clique in

there are 1 additional cliques in , that together cover all the vertices in all the rows in which

has a representative. It will be shown below that . Therefore, if then


¯ . On the other hand, if then by simply dividing the number of vertices in

by the size of the largest clique in , we obtain the claimed lower-bound on ¯ .

The Structure of

The graph is an -partite graph; each row of corresponds to a distinct row of . A row of

consists of vertices (arranged in columns, denoted 0 1). We consider a mapping (to

be specified below) of the vertices in each row of to the vertices in the corresponding row of .

We call the vertex of , to which a vertex of is mapped, the image vertex of .

We now describe how the edge set of is constructed. For every edge in , we have an

edge in where and are the images of and respectively. We refer to such an edge

as a direct edge.

In addition, the edge set of is extended to include all the rotations of the above edges as

follows: If contains an edge connecting the th vertex in row to the th vertex in row , then

we also add to edges connecting the mod th vertex in row to the mod th

vertex in row for every 1 1 .

In order to fully describe , it now only remains to describe the image function, which maps

each vertex of to a unique vertex in . However, from the above description alone we can

conclude the following:

Lemma 55 If then ¯ .

Proof: The graph is symmetric under rotation; i.e., let be the graph that results by

rotating all the rows of by columns to the right (in a wraparound manner), then for every

0 , . Now, given a clique of size in , consider the set which consists

of the images in of the vertices in . Clearly, is a clique of , and due to the symmetry of

under rotation, has 1 rotational images that together cover all the vertices of .

The Image Function

For an integer 0 we denote by the set 0 1 . As a preliminary for the description

of the image function, from to , we start off with a simple lemma, showing the existence


of an injection , of some special structure, from any set of size into the range , for some1 .

Lemma 56 For every positive integer , there exists a function : where Ω 5 ,

such that for every distinctmultiset 1 2 3 , 1 2 3 , the sum 1 2 3 mod

as well as its subsums (that is, 1 2 mod and 1 ) are distinct.

Proof: We use an inductive argument to show the existence of a mapping which satisfies

the property that for every distinct multiset 1 2 3 , 1 2 3 , the sum 1 2

3 mod is distinct. Suppose that 0 1 have already been selected from such

that for 0 1 2 3 1, all the sums 1 2 3 mod are distinct. An element

cannot be chosen for if and only if there are 0 1 2 3 4 5 1 such that

1 2 3 4 5 mod

Therefore, at most 5 elements are ineligible at any step and if Ω 5 , then the process can

be carried through to yield the desired mapping.

Now observe that the mapping shown to exist above must also satisfy the property that for

every distinct multiset of size two, namely 1 2 , the sum 1 2 mod is distinct. This

follows because if there exists two distinct mulisets of size two, namely 1 2 and 1 2 such

that

1 2 1 2 mod

then we can simply add the same element to both multisets and get a contradiction to the property

proven above that the sums of all triplets are distinct. It is clearly the case that the mapping is an

injection.


The Reduction

Let be the injection shown to exist in Lemma 56 for the domain of all vertices of , i.e., ,

and let , the size of the range of . Hence,

: 0 1 0 1

The image of a vertex in the th row of is the vertex of which is in the th row and

column of .

We now show that our transformation of graph into the graph preserves the clique number,

that is, and have the same clique number.

Lemma 57 .

Note that this immediately yields our goal concerning the chromatic number of ¯ since if

then ¯ by Lemma 55. Otherwise, and then simply counting the number of

vertices of and dividing by the size of the largest clique in (using Lemma 57), we get

¯

Proof (of Lemma 57):

: The image of a clique in forms a clique in .

: We say that an edge in is the origin of an edge in if is obtained

via the images of and or by rotating the edge so obtained.

Using the special property of the image function , we first show that every edge in has a

unique origin in . Suppose, to the contrary, that an edge in has two origins in , namely the

edges 1 1 and 2 2 . Then we must have

1 1 2 2 mod

and it follows that

1 2 2 1 mod


which contradicts the property of .

Given a clique in , consider the origins of the edges connecting the vertices of between

themselves. We claim that these origin edges are related to consistent vertex origins in the graph

in the following sense: there exists a mapping , from every vertex , to a vertex ,

such that the origin of the edge 1 2 , for 1 2 , is the edge 1 2 in . Once this

is proven, it follows immediately that the consistent vertex origins of form a clique in , which

implies our claim.

Assume, by way of contradiction, that the origins of the edges connecting vertices in are

not consistent; then there would have been a triangle such that the origin of the edge

is 1 and the origin of the edge is 2 , where 1 2. We show that such

a triangle cannot exist in .

Let, 1 1 be the origin in of the edge , and observe that:

1 col col mod

1 1 col col mod

2 col col mod

where col , for a vertex of , denotes the column of in . Combining the above equivalence

relationships, we get

1 1 1 2 0 mod

and therefore

1 1 1 2 mod

which by the special property of , and the fact that the origins of the edges connecting vertices ,

and are in three distinct parts of , implies that 1 2 (as well as that 1 and 1 ).

Hence we conclude that .


6.3 Hardness of Approximating the Chromatic Number when the Value

is (a Large) Constant

We now sketch how the argument in the previous section can be extended to eliminate the dependency

of on and thereby show that for any constant there exists a constant such that it is NP-hard

to determine whether or .

We start with the graph as described earlier, however, we now assume that is a constant.

Furthermore, the injection which was previously defined over all the vertices in , is now defined

only over the vertices in any part of , that is, the domain of is restricted to . As a result,

several vertices may now map to the same column in and thus an edge in may not have a

unique origin.

We will use the mapping to transform into a -partite graph for some integer

1 , such that if then ¯ , and otherwise ¯ 2 . This transformation

is better described through an intermediate graph where the graph is also a -partite

graph. For every row of , we include a block of rows in such that the th row in the

block corresponding to the row of is simply the th row shifted by columns to the right in a

wraparound manner. Thus each vertex of has copies in the graph . For every edge in

, we insert an edge between every copy of vertex and every copy of vertex in . While doing

so, we assume that each vertex of is connected to itself. Thus all the copies of any vertex form

a clique in . It is easy to see that .

Once the graph is constructed, we transform it into the graph in similar manner as before

by applying the mapping to the vertices in each row of . The edge set of is also constructed

in a similar manner. For every edge in , we have an edge in where and are

the images of and respectively. We extend this edge set by including all the rotations of these

edges. Let be the th vertex in the th row of , and let be the vertex which is the copy of in

the th row of the th block of , then the image of is the vertex of which is in the th row of

the th block and mod column of .

As noted earlier, a consequence of the fact that the domain of is now restricted to is that

an edge in may now have multiple origins in . However, the following claim shows


that this can happen only when and are in the same column in .

Claim 1 Every edge in such that has a unique origin edge in .

Proof: Consider an edge in which has at least two distinct origins edge and let

1 1 and 2 2 be any two such distinct origin edges. Thus we must have

1 1 2 2 mod

By the special property of , we must have 1 2 2 1 . But since 1 1 and 2 2 are

distinct edges, either 1 2 or 1 2. So it must be the case that 1 1 and 2 2. This

immediately implies 1 1 .

We will now relate the clique numbers of the graphs and . Unlike the previous section where

we obtained matching upper and lower bounds on in terms of , we now characterize a

range of values where may lie.

Lemma 58 .

This characterization immediately yields our goal concerning the chromatic number of ¯ since, if

, then contains a clique of size which along with its 1 rotational images

covers all the vertices in and thus ¯ . Otherwise, and therefore,

By choosing , we get 2 . Now simply dividing the total number of vertices in

by the size of the largest clique in , we get ¯ 2 .

Proof (of Lemma 58):

: The image of a clique in forms a clique in and .

: To see this, consider any clique in . Ignore any blocks in

where the clique contains at most one representative. Thus we restrict ourselves to a subset

such that in every block of , either has zero or at least two representatives. Clearly,

because can have precisely one representative in at most blocks. Let denote


the number of blocks in such that contains at least two representatives in them. We will show

that .

Consider now an edge connecting two vertices such that and are in the same

block of . We claim that it must be the case that these two vertices are in different columns and

thus the edge has a unique origin in . This follows rather easily from the observation that

for every edge in such that and are in the same block (and thus they are the copies of

the same vertex in ) . We further use this observation to define a labeling for

each edge where are in the same block in . We define if the edge

is the origin of the edge and correspond to two different copies of the vertex

in .

We now show that for any edge 1 1 1 within a block of , and an edge 2 2 2

also within a block of , where 1 2 1 2 form a 4-clique in , it must be the case that 1

is connected to 2 in . Therefore, the labels of the edges connecting vertices (within the same

block) in form a clique in , which implies our claim that .

Since 1 and 1 are in different columns, as well as 2 and 2, we can assume, without loss of

generality, that 1 and 2 are in different columns (as well as 1 and 2). Hence, the origin in

of the edge 1 2 in is uniquely defined, say 1 2 ; we show that 1 is a copy of 1 and

2 is a copy of 2 in , which are connected in , hence 1 is connected to 2 in .

Let 1 1 be the origin in of the edge 1 1 , we show that it must be the case that

1 1. A similar argument shows that the origin of 2 is consistent with the origin of 1 2 .

We break the argument into two cases:

1 and 2 are in different columns. The same argument as the triangle argument from the

last section applies.

1 and 2 are in the same column. It must be the case that

1 1 1 2 mod

Using the special property of , we know that 1 2 1 1 . Since 1 2 , it


must be the case that 1 1.

6.4 Hardness ofApproximating theChromaticNumber for SmallValues

In this section we describe our proof that distinguishing between the case that the chromatic number

of a graph is at most 3 and the case that it is at least 5 is NP-hard.

The Starting Point

As before, we start with an -partite graph that either has an -clique or every clique of

contains less than 2 vertices. Separating between these two cases is NP-hard [37, 8].

Our Approach

Given such a graph , we construct a graph such that ¯ 3 when and ¯ 5

when 2 . Since the chromatic number of the graph ¯ is equal to the clique cover number

of , our result follows from the construction of such a graph .

For the purpose of this reduction, we assume that every vertex of is connected to itself, and

that every vertex is connected to at least one vertex in each row of . It is easy to see that given the

graph , it can always be transformed into a graph such that satisfies both these assumptions

and iff and 2 otherwise.

6.4.1 The Structure of

The graph is a multi-partite graph; the rows of (partite sets) are divided into blocks, one for

each row (partite set) of . The th block of consists of rows (more precisely, 5 7

rows), where is the number of vertices in the th row of . Therefore, the number of rows in

is at most , where is the maximum number of vertices in any row of .

Each of the rows in the th block of consists of 3 vertices (arranged in 3 columns, denoted

0, 1 and 2) and is associated with a ordered 3-way partition of the set of vertices of the th row of


(however, not all possible such partitions will be used in constructing the graph ). Thus each

vertex in a row in the th block of is labeled by a single subset of the vertices of the th row of

, which we refer to as the label set of that vertex. These labels, as we will see next, determine the

edge set of .

A vertex of , labeled by a set of vertices , is connected to a vertex labeled by the set

, if there exists a vertex and a vertex such that and are connected in .

(Note that due to our assumption that each vertex of is connected to itself, two vertices in the

same block of whose label sets have a non-empty intersection are connected in .) The edges

inserted in this manner are referred to as direct edges.

In addition, the edge set of is extended to include all the rotations of the above edges as

follows: If contains a direct edge connecting the th vertex in row to the th vertex in row ,

then we also add to edges connecting the 1 mod 3 th vertex in row to the 1 mod 3 th

vertex in row , and similarly between the 2 mod 3 th and 2 mod 3 th vertices in these

rows.

For conciseness, a row whose vertices 0, 1 and 2 have labels , and respectively, is said

to have a row label of the form:

In order to fully describe , it now only remains to describe the ordered partitions associated

with each of its rows. However, from the above description alone we can conclude the following.

Claim 2 Let be a subset of the rows of , and let 0 be a clique that has one representative in

each row in . Then the set of rows is coverable by 3 cliques.

Proof: The graph is symmetric under rotation; i.e., let be the graph that results by rotating

all rows of by 1 to the right (in a wraparound manner), then . Therefore,

0 has two rotational images, 1 and 2, that together cover all vertices in all rows in .


The Ordered Partitions

Let 1 denote the set of vertices in the th row of , and for each 1 1 , let

1 , 1 and 1 ( . We now describe how

the th block of rows in is constructed:

The first row is the trivial partition putting all the vertices in one set:

For each , 1 , contains one row with labeling given by

and another row that corresponds to the following ordering of the same partition :

In addition, for each , 1 , contains two rows whose labeling corresponds to a

partition that singles out the unique vertex which is not included in :

and


And finally, for each , 1 , contains a single row with labeling

1

6.4.2 A Clique of Size in Implies a 3-Coloring of ¯

The following lemma is a rather immediate consequence of the construction of the edge set of

and claim 2.

Lemma 59 If then ¯ 3.

Proof: Let 1 be the set of vertices forming a clique of size in . Clearly,

each row of contains one vertex whose labeling contains some ; let 0 be the set of those

vertices in . By the construction of the edge set of , 0 constitutes a clique with a representative

in every row of . Using Claim 2, we conclude that is coverable by 3 cliques.

6.4.3 A 4-Coloring of ¯ Implies a Large Clique in

Our objective now is to show that ¯ 4 implies . We will show that if ¯ 4

then 2 , which, by the constraint on the values taken by , implies that must be

.1

Assume has a 4-clique cover. We use such a cover to identify a set of vertices of , say

1 , where each belongs to a distinct row of such that is a union of two

cliques 1 and 2 in . The construction of such a set is clearly sufficient to conclude that

2 .

The Critical Cliques

Consider the th block of rows in . There are three cliques that contain a vertex in the row whose

row label is of the form

1As an aside, it may be noted that by Lemma 59, this in turn implies that ¯ 3.


where denotes the set of vertices in the th row of . To each of these cliques we assign a shift,

which is either 0,1 or 2, according to whether the clique contains the first, second or the third vertex

respectively.

Observe that the vertices in two rows (in different blocks) with row labels of the above form,

are connected only if they appear in the same column. Hence a clique is assigned in this manner at

most one shift value over all the blocks in . However, one of the shift values may be assigned to

two cliques. Nevertheless, we see that in such a case, we can replace all occurrences of one of the

two cliques by the other in all rows of the above form. This follows from the structure of and

the assumption that any vertex in is connected to at least one vertex in any row of .

Therefore, we can assume from now on that all the rows of the above form are covered by

precisely three cliques in the 4-clique cover of . Let 0, 1 and 2 denote these three cliques

such that is assigned shift value in every block. We refer to these three cliques as the critical

cliques while the remaining clique (which may be empty), denoted by , is referred to as the

non-critical clique.

The Voting Scheme

In any row of , whose label is of the form

0 1 2

we say that the critical clique votes for if contains the vertex with label where

mod 3. That is, 0 votes for the set labeling the vertex it contains, while 1 and 2 vote

for the set labeling the vertex to the immediate left and right (in a wraparound manner) respectively,


of the vertex they contain in the above row. Thus in a row whose label has the form

each of 0, 1 and 2 vote for .

Let us now consider a pair of rows in the same block of which have the form

0 1

1 0

and consider the votes that may be casted by the critical cliques in that pair. The following claims

summarize some useful observations.

Claim 3 In a row whose row label is of the form,

0 1

no critical clique can vote for the empty set.

Proof : By definition, the critical clique appears in the column of the row whose row label is

of the form

Therefore, by our construction of the edge set of , can only appear in the columns and

1 mod 3 of the given row. Thus it can only vote for either the entry in column 0 or the entry

in column 1 of the given row. The claim follows.

Claim 4 In a pair of rows of of the form


0 1

1 0

the following two properties are always satisfied :

(a) a critical clique which appears in both rows either votes for 0 in both rows or votes for 1

in both rows, and

(b) two critical cliques which appear in both rows, either together vote for 0 or together vote

for 1.

Proof : By Claim 3, we know that a critical clique only votes for 0 or 1 in either of the two

rows. Since there are no vertical edges between the two rows, it cannot vote for 0 in one row and

1 in the other row of the pair. Thus property (a) follows. To see property (b), consider a critical

clique, say , which appears in both rows. By property (a), it either votes for 0 or 1 in both

rows. Without loss of generality, assume it votes for 0 in both rows. Observe now that the vertex

which corresponds to the critical clique 1 mod3 voting for 1 in the second row of the pair, is

taken by . Similarly, the vertex which corresponds to 2 mod3 voting for 1 in the first row of

the pair, is also taken by . Thus if either 1 mod3 or 2 mod3 also appears in both rows of

the pair, it must also vote for 0.

We say that the label is elected in the above pair of rows if the majority (i.e. two) of the

critical cliques vote for , 0 1 . The following is a straightforward consequence of the

preceding two claims.

Claim 5 In a pair of rows of of the form

0 1

1 0

majority is always defined.


Proof : If all three critical cliques appear in this pair of rows, majority is clearly defined by

Claim 4(a). On the other hand, if only two of the critical cliques appear in this pair of rows (i.e.,

the pair of rows is covered by only three cliques), then each of these two critical cliques appears in

both the rows and therefore by Claim 4(b), both of them must vote for either 0 or 1.

For example, in Figure 6.1, 0 votes for 0 while 1 and 2 vote for 1. Hence 1 is elected

in this pair of rows.

0

0

1

1

1 0

0 2

Figure 6.1: An example of critical cliques choosing 1

Singling Out the Singletons

We have established so far that in every pair of rows which corresponds to two different orderings

of a partition of the form 0 1 , either 0 or 1 must get majority of the votes. Our next goal

is to show that among all such pairs of rows in a block, there exists one such that the set elected is

a singleton.

Lemma 60 In every block of rows in , there exists a pair of rows corresponding to the partition

of the form where 1 is the set of vertices in the th row of ,

such that is elected by the majority of the critical cliques in this pair.


Proof: Recall that we defined 1 , 1 and 1,

where 1 1 . Now consider the sequence of pairs of rows with the row labels of the form:

If 1 1 is elected in the first pair ( 1), or 1 is elected in the last pair ( 1),

we are done. Otherwise, there must be a switch point 2 1 such that 1( )

is elected in the 1 th pair and ( 1 ) is elected in the th pair. We show that then

it must be the case that is elected in the pair of rows with row label of the form:

Suppose by way of contradiction, the vertex with label is elected in the above pair of rows; we

show that this makes it impossible to cover all three vertices in the row whose row label is:

1

We argue that at most one of the critical cliques can cover a vertex in ; this clearly suffices as

there is only one non-critical clique.

Consider the three pairs of rows which corresponds to the following three partitions:

The partition just before the switch: 1 1 ,

The partition just after the switch: , and

The partition which singles out the vertex that causes the switch: .

A critical clique that contains a vertex in row , votes for one of the three sets, 1 or

. Let denote this set. Clearly, one of the above three partitions has the form .


Let be the pair of rows corresponding to this partition. By our assumption, the label is

elected in . The critical clique cannot vote for in since votes for in and the

edges connecting the vote for in to the votes for in do not exist in (there are three

direct edges connecting a row in to , however, none is a rotation of the edge that if existed in

were to connect in to in ). Therefore, it must be the case that the two remaining

critical cliques, namely 1 mod3 and 2 mod3, vote for in (otherwise would

not have been elected in ).

Let 1 2 0 1 be the columns, of the vertices in thefirst and second rows of respectively,

that are labeled by . Consider the vertex in the first row and the 1 mod 3 column of

, and the vertex in the second row and 2 mod 3 column of (if the critical clique were

to vote for , it would cover at least one of these two vertices). These two vertices cannot be

contained in the critical cliques 1 mod3 or 2 mod3 as this would mean these cliques do not

vote for , which then could not have been elected. Therefore, these two vertices must be

contained in the remaining non-critical clique (see figure 2). Thus contains vertices in both

column and column 1 mod 3 (this follows because 1 2 0 1 ).

Now if another critical clique, say ( ), appears in row voting for some set ,

there exists a different pair of rows, say , such that it corresponds to a partition of the form

and by our assumption, the label is elected in . By applying the same

argument as before, we can conclude that must contain a vertex in column and a vertex in

column 1 mod 3 of this pair. This means contains vertices in all three columns in the

pairs and . However, taking into account the edges connecting pairs and , this contradicts

the following simple claim:

Claim 6 In any block of , consider a pair of rows, say 1, with labels of the form:

0 1

1 0


and a clique in that contains vertex , for 0 1 2 , in the first row and vertex 1 mod 3

in the second row (i.e., some shift of a choice that corresponds to label 0 ). Now consider the

pair of rows, say 2, with row labels of the form :

0 1

1 0

Then cannot contain in this pair of rows a vertex in the 2 mod 3 column.

Proof: Simply observe that the vertex in the first row of 1, is not connected to the vertex

2 mod 3 in the first row of 2 and similarly, the vertex 1 mod 3 in the second row of 1,

is not connected to the vertex 2 mod 3 in the second row of 2.

Consequently, only one critical clique can appear in and thus all the vertices of could not

have been covered by this clique cover. This is a contradiction. We therefore conclude that the

vertex with label is elected in the pair of rows corresponding to the partition .

The Singletons Form at most Two Cliques in

Our final goal now is to show that the set of vertices formed by the singleton sets elected in each

block (we associate exactly one singleton set with each block), is indeed the set we described

earlier.

Lemma 61 Let 1 be a set of vertices of , such that belongs to the th row of

, and is elected in the th block of ; then is a union of two cliques in .

Proof: Let denote the set of vertices in the th row of , where 1 and let denote

a pair of rows of the form below :


1 1

1 1

1

0 1

Figure 6.2: Critical cliques 0 and 1 vote for 1 and , respectively, in row . This forces thenon-critical clique to cover 4 vertices, as indicated, which do not induce a clique in .


where . We say that holds with shift 0 1 2 in if contains both the

vertex in the 1 mod 3 th column of the first row, and the vertex in the th column in the second

row.

We define 1 to be the set of vertices such that, in the pair of rows , holds

with some shift 0 1 2 and let 2 1. We claim that 1 and 2 each form a

clique in .

Let us first look at the easier case which is that of 2. In each , such that 2, does

not hold with any shift . Therefore, every critical clique that contains a vertex in votes for

and hence it must be the case that one of the critical cliques votes for in and in

, and contains 3 vertices in these four rows. This is not possible unless is connected in to

.

Now, regarding 1, we are given such that 1, and such that holds with

shift in and it holds with shift in . We show that, unless is connected in to ,

a 4-clique-cover, in which both and are elected, is not possible.

For clarity of exposition, let us represent the pair of rows and such that their first row is

shifted one column to the left (note that we are not changing the edge set of ). Hence, we have

that looks like:

and that looks like:

The following claim is immediate now.

Claim 7 Unless is connected to in , has no vertical edges connecting the first (second)

row of to the second (first) row of .


By our assumption, holds with some shift in and holds with some shift in

, i.e., contains two vertices in the same column in both and . If , by claim 7 we

are done. Hence we assume, from now on, that ; let 0 1 2 be such that and

. Note that does not contain any vertices in column .

It must be the case that the critical cliques with shifts and constitute the majority of votes

in , and the critical cliques with shifts and constitute the majority of votes in .

Since votes for in and in , contains vertices only in column .

By claim 7, unless is connected to , either contains vertices only from the first row of

and the first row of or the second row of and second row of . However, clique in

and clique in can contain either the vertex in the first row in column in and the second

row in column in , respectively, or vice versa. This yields a contradiction.

6.4.4 Off with the 4th Clique

We can now conclude the following theorem:

Theorem 29 Coloring a 3-colorable graph by 4 colors is NP-hard.

Proof: From Lemma 61 we conclude that if ¯ 4 then 2 , therefore .

The following is a rather straightforward corollary of theorem 29:

Corollary 10 For any fixed , it is NP-Hard to color a -chromatic graph with at most 2 3 1

colors.

Chapter 7

Conclusion

154

CHAPTER 7. CONCLUSION 155

The last three decades have witnessed a continued progress in the design of approximation algo-

rithms with provable performance guarantees. This progress has been complemented by a series

of recent breakthroughs in hardness of approximation which have provided almost matching inap-

proximability results for many important optimization problems. Meanwhile, algorithmic research

in approximation has witnessed another phenomenon whereby problem-specific techniques were

often distilled into algorithmic paradigms which apply to entire classes of problems. The hard-

ness of approximation research, on the other hand, has not seen a similar success in translating

problem-specific results into general principles which uniformly apply to classes of problems.

Our work has developed frameworks whereby the hardness of approximation results for individ-

ual optimization problems could be effectively translated into a structural understanding of an entire

class of optimization problems. For instance, our structure theorem translates the recently obtained

hardness of approximation results for problems such as MAX 3-SAT and MAX CLIQUE into a

structural characterization of approximation classes APX and poly-APX, respectively. Similarly,

our study of constraint satisfaction-based maximization problems identified uniform transformations

to show an approximation-preserving equivalence between well-understood individual problems on

the one hand, and entire classes of problems on the other hand. These results represent new insights

towards understanding the structure of the above-mentioned classes.

But this research represents only a relatively small step towards understanding the approximation

behavior of optimization problems. Our work highlights a number of issues which deserve further

investigation. For example, while the structure theorem yields a complete characterization for many

approximation classes, the class PTAS still seems far from being well-understood. An important

open problem is to identify natural complete problems for this class. Since the running time of a

PTAS problem may have an arbitrary dependence on the error, it seems that any such completeness

results would require reductions which impose weaker restrictions than the -reductions. In

particular, the PTAS reductions of Crescenzi and Trevisan [29] which allow the running time of

reductions to depend on the error, may be an appropriate choice for this task.

Schaefer’s framework of constraint satisfaction problems gave us a way to define two natural

classes of maximization problems. Maximization problems built in this manner represent a well-

behaved microcosm of NPO which captures many representative optimization problems. The

CHAPTER 7. CONCLUSION 156

amenability of these classes to a unified analysis, in turn, provided us with a useful instrument

to gain some formal insight into the approximation behavior of naturally-arising NP maximization

problems. A natural complement to this work would be a study of minimization problems defined on

a similar platform. Khanna, Sudan, and Trevisan [72] have initiated such a study and have obtained

classification theorems which help unify many unresolved problems concerning the approximability

of minimization problems. An interesting new line of research is to discover other ways of defining

natural classes of optimization problems which are amenable to a formal analysis with respect to

their approximation properties. While a collective body of formal results over many such well-

behaved sub-classes of NPO may still not lead to a comprehensive understanding of the NPO as a

whole, it could provide us with a much better understanding of NPO problems that arise in practice.

An alternate direction for extending the above work is to consider constraints over non-boolean

domains.

Finally, semidefinite programming based relaxations have resulted in improved performance

guarantees for 3-coloring [65]. But these guarantees still require a polynomial number of colors for

a 3-colorable graph. The huge gap between the upper and the lower bounds on the approximability

remains virtually untouched by the modest progress described in this work. A resolution of this gap

is perhaps one of the most fundamental open problems in approximation today. Any progress in

merely changing the order of magnitude of this gap would represent a big step in our understanding

of this problem.

Appendix A

Problem Definitions

A.1 SAT

INPUT : A collection of disjunctive clauses of literals over a set of variables .

QUESTION : Is there a truth assignment to which satisfies all clauses in ?

A.2 -SAT

INPUT : A collection of disjunctive clauses of literals over a set of variables such that each

clause has at most literals.

QUESTION : Is there a truth assignment to which satisfies all clauses in ?

A.3 MAX -SAT

INPUT : A collection of disjunctive clauses of literals over a set of variables such that each

clause has at most literals.

GOAL : Find a truth assignment to which satisfies the largest number of clauses.

157

APPENDIX A. PROBLEM DEFINITIONS 158

A.4 MAX CUT

INPUT : A graph .

GOAL : Find a partition of into disjoint sets and so as to maximize the number of edges in

with one end-point in and the other in .

A.5 - MIN CUT

INPUT : A directed graph and two vertices .

GOAL : Find a partition of into disjoint sets and so as to minimize the number

of edges in going from to .

A.6 MAX CLIQUE

INPUT : A graph .

GOAL : Find a largest size subset such that any two vertices in are connected through

an edge in .

A.7 MAX INDEPENDENT SET

INPUT : A graph .

GOAL : Find a largest size subset such that no two vertices in are connected through an

edge in .

A.8 GRAPH COLORING

INPUT : A graph .

GOAL : Find a coloring that uses minimum number of colors (A map : 1 is a coloring

if for all ).

APPENDIX A. PROBLEM DEFINITIONS 159

A.9 GRAPH -COLORING

INPUT : A -colorable graph .

GOAL : Find a coloring of minimizing the total number of colors.

A.10 TSP(1,2)

INPUT : A complete graph such that each edge in has an assigned length of either

one or two.

GOAL : Find a shortest cycle which visits all vertices in .

A.11 MIN VERTEX COVER

INPUT : A graph .

GOAL : Find a partition of into disjoint sets 1 2 such that each is an independent set

and the total number of such sets is minimized.

A.12 MIN SET COVER

INPUT : A pair where is a finite set and is a collection of subsets of .

GOAL : Minimize such that and .

Bibliography

[1] P. Alimonti. New Local Search Approximation Techniques for Maximum Generalized Satisfiability

Problems. In Proceedings of the 2nd Italian Conference on Algorithms and Complexity (1994), pp. 40–

53.

[2] S. Arora. Probabilistic Checking of Proofs and Hardness of Approximation Problems. PhD Thesis, UC

Berkeley (1994).

[3] S. Arora. Polynomial Time Approximation Scheme for Euclidean TSP and Other Geometric Problems.

In Proceedings of the 37th Annual IEEE Symposium on Foundations of Computer Science (1996). To

appear.

[4] S. Arora, L. Babai, J. Stern and Z. Sweedyk. The Hardness of Approximate Optima in Lattices, Codes,

and Systems of Linear Equations. In Proceedings of the 34th Annual IEEE Symposium on Foundations

of Computer Science (1996), pp. 724–733.

[5] S. Arora, D. Karger, and M. Karpinski. Polynomial Time Approximation Schemes for Dense Instances

of NP-hard Problems. In Proceedings of the 27th Annual ACM Symposium on Theory of Computing

(1995), pp. 284-293.

[6] S. Arora and C. Lund. PolynomialTime ApproximationScheme for Euclidean TSP and Other Geometric

Problems. In Approximation Algorithms for NP-hard Problems (D. Hochbaum, Ed.), PWS Publishing,

Boston (1996).

[7] S. Arora, C. Lund, R. Motwani, M. Sudan and M. Szegedy. Proof Verification and Hardness of

Approximation Problems. In Proceedings of the 33rd Annual IEEE Symposium on Foundations of

Computer Science (1992), pp. 14–23.

[8] S. Arora and S. Safra. Probabilistic Checking of Proofs; a New Characterization of NP. In Proceedings

of the 33rd IEEE Symposium on Foundation of Computer Science (1992), pp. 1–13.

160

BIBLIOGRAPHY 161

[9] G. Ausiello, P. Crescenzi and M. Protasi. Approximate Solution of NP Optimization Problems. Theo-

retical Computer Science, 150 (1995), pp. 1–55.

[10] G. Ausiello and M. Protasi. Local Search, Reducibility, and Approximability of NP Optimization

Problems. Inform. Process. Lett., 54 (1995), pp. 73–79.

[11] L. Babai. Trading group theory for randomness. In Proceedings of the 17th Annual ACM Symposium

on Theory of Computing (1985), pp. 421–429.

[12] L. Babai. Transparent Proofs and Limits to Approximation. Technical Report TR-94-07, Dept. of

Computer Science, University of Chicago (1994).

[13] B.S. Baker. Approximation algorithms for NP-complete problems on planar graphs. Journal of the

ACM, 41:153–180 (1994).

[14] I. Barland, P. G. Kolaitis and M. N. Thakur. Integer Programming as a Framework for Optimization

and Approximability. In Proceedings of the 11th Annual Conference on Structure in Complexity Theory

(1996).

[15] M. Bellare. Proof Checking and Approximation: Towards Tight Results. In Sigact News (Complexity

Theory Column), Vol.27, No. 1 (1996).

[16] M. Bellare, O. Goldreich, and M. Sudan. Free bits, PCPs and non-approximability – towards tight

results. In Proceedings of the 36th Annual IEEE Symposium on Foundations of Computer Science

(1995), pp. 422–431.

[17] M. Bellare, S. Goldwasser, C. Lund, and A. Russell. Efficient probabilistically checkable proofs. In

Proceedings of the Twenty-Fifth Annual ACM Symposium on Theory of Computing (1993), pp. 294–304.

[18] P. Berman and M. Furer. Approximating Maximum Independent Set in Bounded Degree Graphs. In

Proceedings of the Fifth Annual ACM-SIAM Symposium on Discrete Algorithms (1993), pp. 365–371.

[19] A. Blum. An 0 4 approximation algorithm for 3-coloring. In Proceedings of the 21st Annual ACM

Symposium on the Theory of Computing (1989), pp. 535–542.

[20] A. Blum. Some Tools for Approximate 3-Coloring. In Proceedings of the 31st IEEE Symposium on

Foundations of Computer Science (1990), pp. 554–562.

[21] H. L. Bodlaender. Some Classes of Graphs with Bounded Treewidth. Bulletin of the European

Association for Theoretical Computer Science (1988), pp. 116–126.

BIBLIOGRAPHY 162

[22] H. L. Bodlaender. A Linear Time Algorithm for Finding Tree-decompositions of Small Treewidth. In

Proceedings of the Twenty-Fifth Annual ACM Symposium on Theory of Computing (1993), pp. 226–234.

[23] D.P. Bovet and P. Crescenzi. Introduction to the Theory of Complexity. Prentice-Hall, New York

(1993).

[24] N. Chiba, T. Nishizeki, and N. Saito. Applications of the Planar Separator Theorem. Journal of

Information Processing, 4:203–207 (1981).

[25] S. A. Cook. The complexity of theorem proving procedures. In Proceedings of the 3rd Annual ACM

Symposium on Theory of Computing (1971), pp. 151–158.

[26] N. Creignou. A Dichotomy Theorem for Maximum Generalized Satisfiability Problems. Journal of

Computer and System Sciences, 51:3: 511–522 (1995).

[27] N. Creignou and M. Hermann. Complexity of Generalized Satisfiability Problems. Research report

1994-2, Group de recherche en algorithmique et logique, Caen (1994).

[28] P. Crescenzi and A. Panconesi. Completeness in approximation classes. Information and Computation,

vol. 93 (1991), pp. 241–262.

[29] P. Crescenzi and L. Trevisan. On approximation scheme preserving reducibility and its applications.

Lecture Notes in Computer Science, vol. 880 (1994), pp. 330-341.

[30] P. Crescenzi, V. Kann, R. Silvestri, and L. Trevisan. Structure in approximation classes. In Proceedings

of the 1st Annual InternationalConference onComputingandCombinatorics, Lecture Notes in Comput.

Sci. 959, pp. 539–548. Springer-Verlag (1995).

[31] P. Crescenzi, R. Silvestri, and L. Trevisan. To Weight or not to Weight: Where is the Question? In

Proceedings of the 4th Israel Symp. on Theory of Computing and Systems (1996), pp. 68–77.

[32] E. Dahlhaus, D.S. Johnson, C.H. Papadimitriou, P.D. Seymour, and M. Yannakakis, The complexity

of multiway cuts. In Proceedings of the 24th Annual ACM Symposium on Theory of Computing,

pp. 241–251 (1992).

[33] K. Edwards, The complexity of colouring problems, Theoretical Comput. Sci., 43:337–343 (1986).

[34] R. Fagin. Generalized First-Order Spectra and Polynomial-time Recognizable Sets. In Richard Karp

(ed.), Complexity of Computer Computations, AMS (1974).

[35] T. Feder and M. Vardi. Monotone monadic SNP and constraint satisfaction. In Proceedings of the 25th

ACM Symposium on Theory of Computing (1993).

BIBLIOGRAPHY 163

[36] U. Feige. A Threshold of ln for ApproximatingSet Cover. InProceedings of the 28thACM Symposium

on Theory of Computing (1996) , pp. 314–318.

[37] U. Feige, S. Goldwasser, L. Lovasz, S. Safra, and M. Szegedy. Approximating clique is almost NP-

complete. In Proceedings of the 32nd Annual IEEE Symposium on Foundations of Computer Science

(1991), pp. 2–12.

[38] U. Feige and J. Kilian. Zero Knowledge and the Chromatic Number. In Proceedings of the 11th Annual

Conference on Structure in Complexity Theory (1996) , pp. 278–287.

[39] A. Frieze and M. Jerrum, Improved approximation algorithms for MAX -CUT and MAX BISEC-

TION. In Proceedings of the 4th InternationalConference on Integer Programming and Combinatorial

Optimization, Lecture Notes in Computer Science 920, pp. 1–13. Springer-Verlag (1995).

[40] M. R. Garey and D. S. Johnson. The complexity of near-optimal graph coloring. Journal of the ACM,

vol. 23 (1976), pp. 43-49

[41] Michael R. Garey and David S. Johnson. Computers and Intractability: A Guide to the Theory of

NP-Completeness. W. H. Freeman (1979).

[42] N. Garg, H. Saran, and V. Vazirani. Finding separator cuts in planar graphs within twice the optimal.

In Proceedings of the 35th Annual ACM Symposium on Foundations of Computer Science (1994),

pp. 14–23.

[43] N. Garg, V. V. Vazirani, and M. Yannakakis, Approximate max-flow min-(multi)cut theorems and their

applications. In Proceedings of the 25th Annual ACM Symposium on Theory of Computing (1993), pp.

698–707.

[44] M. X. Goemans and D. P. Williamson, A General Approximation Technique for Constrained Forest

Problems. In Proceedings of the 3rd Annual ACM-SIAM Symposium on Discrete Algorithms (1992),

pp. 307–316.

[45] M. X. Goemans and D. P. Williamson, Improved approximation algorithms for maximum cut and

satisfiability problems using semidefinite programming. Journal of the ACM, 42, pp. 1115–1145

(1995).

[46] S. Goldwasser, S. Micali, and C. Rackoff. The knowledge complexity of interactive proof-systems. In

Proceedings of the 17th Annual ACM Symposium on Theory of Computing (1985), pp. 291–304.

[47] R.L. Graham. Bounds for certain multiprocessing anomalies. Bell Systems Technical Journal, 45:1563–

1581 (1966).

BIBLIOGRAPHY 164

[48] M. Grigni, E. Koustsoupias, and C. Papadimitriou. An Approximation Scheme for Planar Graph TSP.

In Proceedings of the 36th Annual ACM Symposium on Foundations of Computer Science (1995),

pp. 640–645.

[49] M. M. Halldorsson. A still better performance guarantee for approximate graph coloring. Inform.

Process. Lett., vol. 45 (1993), pp. 19–23

[50] M. M. Halldorsson and J. Radhakrishnan. Improved Approximations of Independent Sets in Bounded-

Degree Graphs. In Proceedings of the 4th Scandanavian Workshop on Algorithm Theory (1994),

pp. 194–206.

[51] J. Hartmanis and L. Berman. On isomorphisms and density of NP and other complete sets. InProceedings

of the 8th ACM Symp. on Theory of Computing (1976), pp. 30–40.

[52] J. Hastad. Clique is Hard to Approximate within 1 . In Proceedings of the 37th Annual IEEE

Symposium on Foundations of Computer Science (1996). To appear.

[53] Dorit S. Hochbaum. Efficient bounds for the stable set, vertex cover, and set packing problems. Discrete

Applied Mathematics, vol. 6 (1982), pp. 243–254.

[54] H.B. Hunt III, M.V. Marathe, V. Radhakrishnan, S.S. Ravi, D.J. Rosenkrantz, and R. E. Stearns.

Approximation Schemes using L-reductions. In Proceedings of the Conference on Foundations of

Software Technology and Theoretical Computer Science (1994), pp. 342-353.

[55] N. Immerman. Descriptive Complexity: a Logician’s Approach to Computation. In Notices of the

American Mathematical Society, 42(10): 1127–1133 (1995).

[56] N. Karmarkar and R.M. Karp. An Efficient Approximation Scheme For The One-Dimensional Bin

Packing Problem. In Proceedings of the 23rd Annual IEEE Symposium on Foundations of Computer

Science (1982), pp. 312–320.

[57] R.M. Karp. Reducibility among combinatorial problems. In Complexity of Computer Computations

(R.E. Miller and J.W. Thatcher, Eds.), pp. 85–103, Plenum, New York (1972).

[58] D. S. Johnson. Approximation Algorithms for Combinatorial Problems. Journal of Computer and

System Sciences, vol. 9 (1974), pp. 256-278.

[59] D. S. Johnson. The Tale of the Second Prover. In the NP-completeness Column: An On-going Guide.

Journal of Algorithms, vol. 13 (1992), pp. 502–524.

BIBLIOGRAPHY 165

[60] D. S. Johnson, C. Papadimitriou, and M. Yannakakis. How Easy is Local Search? Journal of Computer

and System Sciences, vol. 37 (1988), pp. 79–100.

[61] V. Kann. On the Approximability of NP-complete Optimization Problems. Ph.D. Thesis, Department

of Numerical Analysis and Computing Science. Royal Institute of Technology, Stockholm, Sweden

(1992).

[62] V. Kann, S. Khanna, A. Panconesi, and J. Lagergren. On the Hardness of Approximating MAX -CUT

and its Dual. In Proceedings of the 4th Israel Symp. on Theory of Computing and Systems (1996),

pp. 61–67.

[63] R. M. Karp. Reducibilityamong combinatorial problems. In Raymond E. Miller and James W. Thatcher,

editors, Complexity of Computer Computations, pp. 85–103. Plenum Press (1972).

[64] D. Karger, R. Motwani and G.D.S. Ramkumar. On approximating the longest path in a graph. In

Proceedings of the Third Workshop on Algorithms and Data Structures (1993), pp. 421–432.

[65] D. Karger, R. Motwani and M. Sudan. Approximate Graph Coloring by Semidefinite Programming. In

Proceedings of the 35thAnnual IEEE SymposiumonFoundationsof Computer Science (1994), pp. 2–13.

[66] S. Khanna, N. Linial, and S. Safra. On the Hardness of Approximating the Chromatic Number. In

Proceedings of the 2nd Israel Symp. on Theory of Computing and Systems (1993), pp. 250–260.

[67] S. Khanna, R. Motwani, M. Sudan, and U.V. Vazirani. On Syntactic versus Computational Views of

Approximability. In Proceedings of the 35th Annual IEEE Symposium on Foundations of Computer

Science (1994), pp. 819–830.

[68] S. Khanna and R. Motwani. Towards a Syntactic Characterization of PTAS. In Proceedings of the 28th

ACM Symposium on Theory of Computing (1996) , pp. 329–337.

[69] S. Khanna, R. Motwani, and S. Vishwanathan. Approximating MAX SNP Problems via Semi-Definite

Programming. In preparation (1996).

[70] S. Khanna and M. Sudan. The Optimization Complexity of Constraint Satisfaction Problems. Stanford

University Technical Note, STAN-CS-TN-96-29 (1996).

[71] S. Khanna, M. Sudan and D.P. Williamson. The Optimization Complexity of Structure Maximization

Problems. In preparation (1996).

[72] S. Khanna, M. Sudan and L. Trevisan. Constraint Satisfaction: The Approximability of Minimization

Problems. In preparation (1996).

BIBLIOGRAPHY 166

[73] Phokion G. Kolaitis and Madhukar N. Thakur. Approximation Properties of NP Minimization Classes.

JCSS, 50(3):391–411 (1995).

[74] Phokion G. Kolaitis and Madhukar N. Thakur. Logical Definability of NP Optimization Problems.

Information and Computation, 115(2):321–353 (1994).

[75] R. Ladner. On the structure of polynomial time reducibility. Journal of the ACM, 22:1, pp. 155–171

(1975).

[76] L. A. Levin. Universal sorting problems. Problems of InformationTransmission, 9, pp. 265–266 (1973).

[77] D. Lichtenstein. Planar Formulae and their Uses. SIAM Journal on Computing, 11:329–343 (1980).

[78] N. Linial and U. Vazirani. Graph Products and Chromatic Numbers. In Proceedings of the 30th IEEE

Symposium on Foundations of Computer Science (1989), pp. 124–128.

[79] R.J. Lipton and R.E. Tarjan. A separator theorem for planar graphs. SIAM Journal of Applied Mathe-

matics, 36:177–189 (1979).

[80] R.J. Lipton and R.E. Tarjan. Applications of Planar Separator Theorem. SIAM Journal on Computing,

9:615–627 (1980).

[81] C. Lund and M. Yannakakis. The approximation of maximum subgraph problems. 20th International

Colloqium on Automata, Languages and Programming, pp. 40–51 (1993).

[82] C. Lund and M. Yannakakis. On the hardness of approximating minimization problems. Journal of the

ACM, 41:960-981 (1994).

[83] R. Motwani. Lecture Notes on Approximation Algorithms. Report No. STAN-CS-92-1435, Department

of Computer Science, Stanford University (1992).

[84] R. Motwani and P. Raghavan. Randomized Algorithms. Cambridge University Press (1995).

[85] A. Panconesi and D. Ranjan. Quantifiers and Approximation. In Proceedings of the Twenty-Second

Annual ACM Symposium on Theory of Computing (1990), pp. 446–456.

[86] C.H. Papadimitriou. Computational Complexity. Addison-Wesley (1994).

[87] C. H. Papadimitriou and M. Yannakakis. Optimization, Approximation, and Complexity Classes.

Journal of Computer and System Sciences, vol. 43 (1991), pp. 425–440.

[88] C. H. Papadimitriou and M. Yannakakis. The traveling salesman problem with distances one and two.

Mathematics of Operations Research, vol. 18 (1993), pp. 1–11.

BIBLIOGRAPHY 167

[89] E. Petrank. The Hardness of Approximation: Gap Location. Computational Complexity, vol. 4 (1994),

pp. 133–157.

[90] S. Phillips and S. Safra. PCP and tighter bounds for approximating MAX-SNP. Manuscript (1992).

[91] S. Poljak and Z. Tuza, The max-cut problem – a survey. In L.Lovasz and P.Seymour editors, Special Year

on Combinatorial Optimization, DIMACS series in Discrete Mathematics and Theoretical Computer

Science, American Mathematical Society (1993).

[92] P. Raghavan and C.D. Thompson. Randomized Rounding. Combinatorica, 7 (1987), pp. 365–374.

[93] N. Robertson and P. D. Seymour. Graph Minors. II. Algorithmic Aspects of Treewidth. Journal of

Algorithms, 7:309–322 (1986).

[94] S. Sahni. Approximate algorithms for the 0/1 knapsack problem. Journal of the ACM, 22:115–124

(1975).

[95] T. Schaefer. The complexity of satisfiability problems In Proceedings of the 10th Annual ACM Sympo-

sium on Theory of Computing (1978), pp. 216–226.

[96] W. Fernandez de la Vega and G.S. Lueker. Bin Packing can be solved within 1 in Linear Time,

Combinatorica, 1:349–355 (1981).

[97] M. Sudan. Efficient Checking of Polynomials and Proofs, and the Hardness of Approximation Problems.

PhD Thesis, UC Berkeley (1992).

[98] A. Wigderson. Improving the performance guarantee for approximate graph coloring. Journal of the

ACM, vol. 30, No. 4 (1983), pp. 729–735

[99] M. Yannakakis. The analysis of local search problems and their heuristics. In Proceedings of the 7th

Annual Symposium of Theoretical Aspects of Computer Science (1990), pp. 298–311.

A STRUCTURAL VIEW OF APPROXIMATIONsanjeev/papers/thesis.pdf · 1 Introduction 1 1.1 The Historical...

Documents

Transcript of A STRUCTURAL VIEW OF APPROXIMATIONsanjeev/papers/thesis.pdf · 1 Introduction 1 1.1 The Historical...